The depth of the honeybee's backup sun-compass systems

Orientation cues that can provide directional information on an earth-wide scale are called compass cues, and there are two general types (Able, 2001; Wiltschko and Wiltschko, 2009): magnetic cues, based on the earth's magnetic field (Wiltschko and Wiltschko, 2005; Walker, 1997; Wajnberg et al., 2010); and celestial cues, based on the stars, sun, moon, and skylight patterns associated with the sun and moon (Wiltschko and Wiltschko, 2009; Wehner, 1994; Wehner, 1997; Wehner and Müller, 2006; Kraft et al., 2011; Ugolini et al., 2009; Warrant and Dacke, 2011; Dacke et al., 2011; Dacke et al., 2013). To use celestial cues as a compass, animals must account for the cues' movements across the sky, for which they use their innate circadian and other physiological rhythms (Wiltschko and Wiltschko, 2003; von Frisch, 1967; Reppert et al., 2010; Ugolini et al., 2007; Meschini et al., 2008). Other cues can provide compass-like directional information more locally, for example the skyline panorama (Graham and Cheng, 2009a; Reid et al., 2011; Philippides et al., 2011), large-scale features of the landscape (Southwick and Buchmann, 1995; Pahl et al., 2011) and the wind (Müller and Wehner, 2007).

Compass orientation appears to be so important that most mobile animals have multiple compass mechanisms (Able, 1991), which can interact in at least three ways. First, they can interact in each others' development, as the magnetic compass of pigeons may serve as a reference against which the celestial rotation is learned (Wiltschko and Wiltschko, 1990; Wiltschko and Wiltschko, 1998; Wiltschko and Wiltschko, 2003), or as celestial cues calibrate the magnetic compass in some migratory birds (Able and Able, 1996; Cochran et al., 2004; Muheim et al., 2009; Liu and Chernetsov, 2012). Second, multiple compasses can act in combination, either simultaneously or over time, making the resultant orientation more flexible and reliable (Able and Able, 1996; Cheng et al., 2007; Muheim et al., 2003). And third, multiple compasses can act as backups to each other, making orientation possible under a wider variety of conditions (Able, 1991; Walcott, 2005; Walraff, 2005; Gould and Gould, 2012).

Honeybees have a magnetic compass (DeJong, 1982; Collett and Baron, 1994; Schmitt and Esch, 1993; Frier et al., 1996; Walker, 1997; Wajnberg et al., 2010), a celestial compass based on the sun and sun-linked patterns of color and polarization in the blue sky (von Frisch, 1967; Wehner, 1994; Wehner, 1997), and a backup celestial compass for cloudy days, the latter based on a memory of the celestial rotation over time in relation to the landscape (Dyer and Gould, 1981; Dyer, 1987). The celestial compasses based on direct views of the sun and skylight polarization patterns might best be considered distinct, as the sun itself and the skylight polarization patterns are detected separately, and the polarization compass dominates when it is made to conflict with direct views of the sun, at least in ants (Wehner and Müller, 2006). Moreover, ants have a third celestial compass mechanism mediated by the ocelli, which appears to specialize in storing information on the most recent leg of a journey (Schwarz et al., 2011). The celestial compasses of bees appear to be the primary ones when they are available, while the magnetic compass serves as a backup in certain contexts (Frier et al., 1996). The magnetic compass may also be the primary compass for comb orientation, although the resultant orientation is weak (DeJong, 1982). Beyond this, the mechanisms and interactions of these compasses are not well understood. We know, for example, that ants (Wehner and Müller, 1993) and bees (Dyer and Dickinson, 1994) have crude innate expectations regarding the celestial movements and that both insects eventually acquire more-or-less accurate knowledge of the celestial rotation throughout the day (Wehner and Lanfranconi, 1981; Dyer, 1987), but we do not know with certainty what directional cue serves as the fixed reference against which this rotation is first learned. We also do not know whether bees have any further backup systems, such as a memory of the celestial rotation in relation to the magnetic field.

In elucidating the interactions of multiple compasses, studies of multiple species under similar conditions and studies of well-known model species under various conditions are both useful (Able, 1991). Here we take the latter approach, challenging a well-known model species to perform sun-compass orientation when all of its known celestial compasses, including its backup compass based on a knowledge of the celestial rotation in relation to the landscape, are unavailable, that is, under overcast skies in unfamiliar landscapes. While bees have no known mechanisms for locating the sun under such conditions, they might plausibly learn the sun's course in relation to the earth's magnetic field, as pigeons seem to do (Wiltschko and Wiltschko, 1990). Bees can use a magnetic compass while learning visual scenes at food sources (Collett and Baron, 1994; Frier et al., 1996), and there is no obvious reason why they could not use their magnetic compass to locate the sun when the landscape fails to be useful.

Our specific strategy, following von Frisch (von Frisch, 1967), Lindauer (Lindauer, 1960; Lindauer, 1971) and others (Edrich et al., 1979; Brines and Gould, 1982; Wehner and Rossel, 1985; Wehner, 1989; Dyer, 1987; Dyer and Dickinson, 1994; Towne and Kirchner, 1998; Towne and Moscrip, 2008), is to exploit the dance communication to gain access to the bees' knowledge of the sun's compass bearing. In their dances, bees indicate the directions toward recently visited resources by performing a series of wagging runs on the vertical combs inside the nest: wagging runs pointing straight upward on the comb indicate a flight in the direction of the sun's azimuth in the field. Dyer used the dances when he first showed that bees can locate the sun on cloudy days by referring to a memory of the sun's compass bearing over time in relation to the landscape (Dyer and Gould, 1981; Dyer, 1987). In a typical experiment, Dyer placed a hive beside an extended landmark such as a treeline and trained bees to visit a feeder placed some distance from the hive along the treeline. After several days, he transplanted the hive to a similar but differently oriented treeline and set out a feeder in its usual place in relation to the treeline, although now in a novel compass direction. On overcast days, such transplanted bees danced as if they were at their home treeline, based on a memory of the sun's course in relation to it. We now know that bees can recall the sun's compass direction in relation to any flight route in the vicinity of their nests, even routes with which they have had little or no specific experience, and probably also in relation to the panoramic silhouette of the skyline around the nest (Towne and Moscrip, 2008). Although the bees' use of the skyline needs further testing, ants clearly orient using the skyline panorama in this way (Graham and Cheng, 2009a).

Dyer (Dyer, 1984) has already performed a single trial of an experiment in which he transplanted a hive between dissimilar sites under cloudy skies and observed the bees' dances on a vertical comb to see whether the bees could locate the sun. They could not: they danced in two main directions roughly opposite each other, both roughly 90 deg from the predicted direction based on the actual locations of the feeder and the (hidden) sun. Dyer concluded that bees probably have no further backup systems for locating the sun but that they may have ‘some shared rule regarding the interpretation of the landmarks’ [p. 114 in Dyer (Dyer, 1984)] – that is, rules by which they might have matched the home and test landscapes – that could account for why many of his bees seemed to agree on incorrect dance directions. Here we report the results of our attempts to repeat Dyer's experiment using three different sets of landscapes in an effort to clarify the bees' backup systems and their rules for matching landscapes.

We worked with bees of mixed subspecific background, mostly Apis mellifera ligustica. Each of the four different colonies was installed in a two-frame observation hive as described in Towne et al. (Towne et al., 2005) and placed at its location in the field several weeks before experiments began to ensure that the bees were native to their respective sites. The natal sites themselves are described further in the Results. We transplanted bees from their natal sites to visually different test sites under overcast skies and then observed the bees' dances as they foraged from feeders placed some distance from the hive. The directions of the dances with respect to the upward vertical on the comb then indicated whether the bees could locate the sun at the test sites.

All of our basic equipment and techniques are described in earlier publications (Towne et al., 2005; Towne, 2008; Towne and Moscrip, 2008). At least several days before the first experiments each year (2009 and 2012), we trained groups of bees from each hive to visit pneumatic feeders offering lightly scented sucrose solutions 1–2 m from the hive entrances. The feeders were generally set out for 2–3 h daily, but the feeding times and durations varied. The bees visiting the feeders were individually marked with numbered tags, although in one trial (20 July 2012) we also marked several newly recruited bees at the test site with small, unique spots of paint on the thorax.

The simplest technique by which we transplanted bees involved covering the hive entrance with a piece of mesh fabric, transporting the hive by car to the recipient site, and re-opening the hive there. A second technique involved capturing approximately 20 bees as they foraged at a feeder, placing them in small fiberglass screen cages, and placing the cages on the top frame of a recipient hive for several hours or overnight. When the caged bees were later released into the hive at the test site, some of the bees typically found the feeder and foraged as full members of their new colony; others were rejected at the new hive. For one experiment (17 and 21 July 2009), the two methods were combined, so that the hive at the test site contained residents of the transplanted hive as well as cage-transplanted bees from another colony. In these cases, the hive itself was transplanted between two treelined sites, and we knew that bees transplanted between these two sites would dance using a memory of their natal treeline when all celestial cues were effectively obscured by clouds (Towne and Moscrip, 2008). This allowed these bees to serve as controls to confirm that there were no celestial cues available to the cage-transplanted experimental bees during the tests. The dances of the treeline-to-treeline control bees did indeed confirm that the bees could not locate the sun during these tests, but because the experimental bees themselves danced as if they could not detect the sun (see Results), we do not present the results for the control bees. Further, we performed some tests without control bees so that we could take advantage of a greater variety of test sites. The minimum transplantation distance used was 1.3 km, which is farther than necessary to prevent the bees from dancing according to a previous familiarity with the test sites during the cloudy-day tests (Towne et al., 2005). An aerial photograph of the entire region showing the relationships of the natal and test sites for all four experiments is shown in supplementary material Fig. S1.

As soon as several bees from a transplanted hive found the feeder at the test site, we began to move the feeder away from the hive. When the feeder was at least 50 m away, the distance at which the dances of these bees normally show obvious orientation, we began recording the dances through the glass window of the hive with a hand-held Canon Powershot SX10 IS (2009) or Nikon D5100 (2012) camera. Several weighted threads hanging over the glass between the camera and the bees indicated vertical for all recordings. The hive and observer were covered with a shroud made of four thicknesses of white fabric, allowing only dim diffuse light to illuminate the dance floor. When this recording arrangement is used on sunny days, the bees' dances are well oriented as predicted by the sun's location (Towne et al., 2005), and when it is used on solidly overcast days with bees that have been transplanted between two treelined sites, the dances are well oriented according to the bees' memories of the sun's course in relation to their natal treeline (Towne, 2008; Towne and Moscrip, 2008). As we recorded dances, we continued to move the feeder away from the hive in the same direction until it reached its final distance, which varied from 75 to 150 m in different tests. During the recordings, an observer at the feeder recorded the sky conditions and the identities of the bees visiting the feeder.

The directions of the bees' dances relative to the vertical threads were estimated from the video recordings to the nearest 7.5 deg based on a visual average of at least five wagging runs within a single bout of dancing. Each bee was scored only once after a single trip to the feeder. For the statistical analyses, we used each bee's mean dance angle for that day as a single observation. That is, each bee is weighted equally in the statistical analyses regardless of how many individual bouts of dancing she performed. Further, all analyses include only clearly oriented dances that occurred under completely overcast skies; disoriented and round dances are reported below but are not included in the statistical analyses, as these gave no single direction. The inter-observer reliability of our measurement method is high: a large sample of sunny-day dances analyzed by two different observers (J.R.K. and W.F.T.) gave a mean absolute difference for individual dances of 5.3 deg (N=119 dances) and an overall difference, with each bee weighted equally, of 1.9 deg (N=20 bees).

Dance directions were analyzed using Oriana circular statistics software (Kovach Computing Services, Anglesey, Wales, UK). When the distributions of angles from two trials of an experiment were not significantly different according to Watson's U² test (Batschelet, 1981), we pooled the results for further analysis. We used the Rayleigh test (Batschelet, 1981) to determine whether a distribution was significantly non-random and the 95% confidence intervals around a mean vector r to determine whether a mean vector was consistent with a hypothetical expectation.

To analyze the possibility that the bees in two experiments oriented according to their memories of the sun's course in relation to the skylines, we compared 360 deg panoramic images of the natal and test sites. Each panoramic image was assembled from 15 individual wide-angle photographs taken with a Canon Powershot SX10 IS camera in PhotoStitch mode using Hugin (an open source graphic user interface for Panorama Tools, hugin.sourceforge.net). The photos were taken from our eye level (1.6 m above the ground) with the camera in portrait orientation, and the camera was rotated around the vertical post of a leveled tripod between images to ensure that the camera was always aimed at the same elevation.

The original panoramic photographs are shown in Fig. 4C for Experiment 1 and Fig. 7 for Experiment 4. The photos were taken at or near the hive at two of the sites (Experiment 1 natal site, Fig. 2A; Experiment 4 test site, Fig. 6B, location D) and 100 m from the hive into the adjacent fields for the other two sites (Experiment 1 test site, Fig. 2B; Experiment 4 natal site, Fig. 6A). For the two photographs taken 100 m from the hives, their locations are indicated by small black + symbols in Fig. 2B and Fig. 6A. These were taken some distance from the hives because treelines near the hives dominated the skylines there as we saw them from ground level, but the bees probably viewed the broader panoramas from well above the ground. Thus we took the photographs far enough from the treelines to show the broader panoramas.

The panoramic photographs were analyzed further using Adobe Photoshop (Adobe Systems, San Jose, CA, USA). To minimize assumptions about how bees process skyline information, we adjusted the resolution of the photographs to 10 pixels deg⁻¹, which greatly exceeds the bees' ~1 deg visual resolution (Land, 1997). We then converted the images to black-and-white silhouettes using Photoshop's threshold function, choosing a threshold that left the entire sky white and non-sky black. The few non-sky objects on the ground that were above the threshold (for example, the white hive and the sky's reflection in a small pond) were blackened manually. We then converted the two silhouettes to different colors, one green and one yellow, and overlaid them in separate layers. Next we set the two skylines' lowest points – their apparent horizons – to the same level and cropped both images just below these points. The resultant composite image from Experiment 1 is shown in Fig. 1.

Next, we attempted to quantify how well the silhouettes matched in different orientations by measuring how much the natal-site and test-site silhouettes overlapped as the two were rotated against each other. In Fig. 1, the silhouette of the natal site from Experiment 1 is shown in green and that of the test site in yellow (indicating overlap with the natal skyline) or red (indicating no overlap). As a simple measure of the correspondence between the silhouettes, we counted the number of pixels for which the silhouettes overlapped (yellow in Fig. 1) for each 5 deg of rotation of the test-site panorama against the other for one complete 360 deg rotation. The amount of overlap was measured by making the test-site layer partially transparent, which made the areas of overlap and non-overlap different in color, and using Photoshop's color-selection and measurement tools to count the pixels in the area of overlap. The results of these measurements show how well the two skylines match in different orientations for comparison with how the bees matched the two sites. The results of the analyses are given below (in Figs 3, 4 for Experiment 1 and Fig. 7 for Experiment 4).

In order to determine whether bees can locate the sun under overcast skies at sites that are visually unlike their natal sites, we transplanted groups of individually marked bees from their natal site in a narrow valley (Fig. 2A) to a sloping treeline beside an agricultural field (the test site, Fig. 2B). We then recorded the dances of these bees as they visited an artificial feeder at the test site under overcast skies. For the first trial, on 17 July 2009, we caged a group of marked foragers at the natal site at 09:30 h (all times are local solar time) and put the cage into a recipient hive, which we then moved to the test site. When the sky became overcast at 13:30 h, we released the caged bees and set out a feeder near the hive entrance. After several of the transplanted bees found the feeder, we moved it slowly up the treeline (white arrow in Fig. 2B) until it reached 125 m from the hive. Meanwhile, as the feeder passed through 50 m (at 16:00 h) we started recording the bees' dances and continued recording, under overcast skies, for the next 2 h. The directions indicated by the bees' dances in this trial are shown in Fig. 2C (gray bars, based on 27 dances by eight bees). Here and throughout, we use the angle of the mean vector for all dances by each bee on a given day in the statistical analyses and figures, regardless of how many dances the bee performed. That is, each bee is represented by a single data value for each trial.

We performed a second trial of this experiment with new foragers 4 days later on 21 July 2009. For this trial, the bees were caged and transplanted into the recipient hive in the late morning on the day before the experiment and released into the hive at the test site at 06:15 h on the test day. We recorded dances from 07:40 to 11:40 h under overcast skies, during which time the feeder was between 60 and 125 m from the hive. The results of this second trial are also shown in Fig. 2C (black bars, based on 40 dances by seven bees).

The distributions of dance directions from the two trials were not significantly different (Watson's U² test, N₁=8 bees, N₂=7 bees, P>0.2), so we pooled the results for further analysis. In Fig. 2C, the dance direction expected for bees that had located the sun correctly is shown by the thin black line (353.5 deg clockwise of N). The dances are significantly oriented (P<0.001, Rayleigh test, N=15), but their mean direction (thick black arrow 182.6 deg clockwise of N; r=0.77) is roughly opposite the correct direction, and only one of the 15 bees danced within 90 deg of the correct direction. Overall, this result is similar to what we see when bees are transplanted between differently oriented twin treelines under clouds: the bees do not locate the sun correctly but instead dance according to a memory of the sun's course in relation to their natal treeline. Here, however, the landmarks at the two sites could hardly have been more different (Fig. 2A,B). What were the bees doing?

It seems likely that the bees matched the skyline panoramas at the two sites and danced according to a memory of the sun's course in relation to their natal skyline. This is exactly what the bees seem to have done in a previous experiment (Towne and Moscrip, 2008), but in the current case the skylines were less well matched. Here the natal site was at the bottom of a narrow valley, while the test site was partway up the slope of a much broader valley. Nonetheless, our analysis of the skylines (Fig. 3) (full explanation in the Materials and methods) indicates that the bees danced according to the best match between the two skylines. Fig. 4 shows the two skyline silhouettes as they match best (panel A) and as the bees matched them (panel B), and panel C shows the full panoramic photographs as the bees matched them. During the tests, the hive was at the test site (Fig. 4C, top), and the bees flew northward toward the feeder (direction labeled F in Fig. 4C, top), but they danced as if they were at their natal site flying southward (direction labeled D in Fig. 4C, bottom). The two panoramas are quite similar in this arrangement if one ignores the trees in the foreground at the natal site (bottom). The bees themselves could separate the foreground from the distant skyline using motion parallax cues (Srinivasan, 2011; Braun et al., 2012). The difference between the best skyline match and the way the bees matched the sites is 10 deg (Fig. 2C; Fig. 3), well within the 95% confidence intervals for the mean dance direction (±23 deg). In this experiment, then, the bees certainly did not locate the sun correctly using some previously unknown backup system; they apparently made the best of the resemblance between the skyline silhouettes at two sites and danced according to that resemblance, albeit with considerable scatter.

In the next experiment, we attempted to find a pair of sites that would foil the bees' skyline-matching mechanism, so we transplanted a colony whose natal site was at a sloping treeline (Fig. 5A) to a test site near the highest point in a dissimilar landscape (Fig. 5B). In the first trial, on 2 August 2009, we closed the hive at the natal site at 04:30 h, moved it to the test site, set out a feeder, and opened the hive at 06:30 h under a completely overcast sky. We moved the feeder toward the south-southwest (white arrow in Fig. 5B) but were interrupted by a period of heavy rain that lasted approximately 80 min. We resumed the experiment at 09:45 h with the feeder 50 m from the hive and recorded 10 dances by seven different bees under a solidly overcast sky for only 15 min until the sun became visible, by which time we had moved the feeder to 75 m from the hive. The mean dance directions of six of the bees are shown in Fig. 5C (black bars). The seventh bee gave only a single disoriented dance that we scored as a round dance (more on disoriented dances below).

We performed a second trial of this experiment 7 days later (9 August 2009), closing the hive at 05:00 h and opening it at the test site at 06:30 h. We started recording dances of marked foragers when the feeder reached 50 m from the hive at 07:52 h and continued moving the feeder until it reached 150 m from the hive. We recorded the dances of 22 different bees under overcast skies with intermittent rain before we ended the experiment 3 h later. The directions indicated by the dances of these 22 bees are shown in Fig. 5C (gray bars). Four of the dancers in this trial had also danced in the first trial, but as these bees had been returned to their natal site immediately after the first trial and a week had intervened between trials, we treated these as new bees for the purpose of the analysis.

Altogether we recorded a total of 168 dances by 29 bees in the two trials of this experiment, although only 28 bees gave directional dances that could be included in the statistical analysis. The directions of the oriented dances from the two trials were not significantly different (Watson's U² test, N₁=6 bees, N₂=22 bees, P>0.2), so we pooled the results, which gives a mean dance direction of 17.6 deg (r=0.39, N=28). This is roughly opposite the direction expected for bees that had located the sun correctly (208.5 deg; thin black line in Fig. 5C). The dance directions are not entirely random (P=0.015, Rayleigh test, N=28), but they are widely scattered, as if the bees were having great difficulty determining which way to dance. Consistent with the latter is the further observation that some of the dances on these days were abnormal. For 16 dances (of 168) performed by seven different bees (of 29), we were unable to identify clear wagging directions. We scored these dances as either round dances or disoriented waggle dances depending on whether they contained clear wagging segments. As these dances gave no clear direction, they are not included in the statistical analysis. Von Frisch described disoriented waggle dances that he observed under similar conditions as follows [p. 395 in von Frisch (von Frisch, 1967)]: ‘It was not possible in the dances on a vertical comb to measure a definite direction, because it varied greatly from one tail-wagging run to another…. Often the dancers turned searchingly round and round.’ Our observations occurred under very similar conditions, and we observed such dances, which we called ‘disoriented waggle dances’. Further, some of the oriented dances of approximately half of the bees in these trials showed one or more abnormalities: (1) the dances were ‘sloppy’, meaning that the bee occasionally waggled in a direction different from that of most wagging segments; (2) the dancer frequently ‘missed turns’, that is, she performed two or more full circles between wagging runs; and (3) the dancers ‘overshot’ the wagging direction, meaning that the bee turned more than 360 deg between wagging runs so that sequential wagging runs aimed alternately to the left and right of the mean direction. Normal dances, by contrast, typically ‘undershoot’ the mean direction (von Frisch, 1967; Towne and Gould, 1988). Dyer (F. C. Dyer, personal communication) observed such overshooting dances in some of his cloudy-day experiments, and Weidenmüller and Seeley (Weidenmüller and Seeley, 1999) sometimes observed such dances in bees dancing on wire screen surfaces to advertise nearby nest sites. We did not attempt to quantify these abnormalities further; they are difficult to score because an individual dance can show any or all of these features to varying degrees at different times.

In summary, the bees in Experiment 2 were transplanted between sites with dissimilar landmarks and skylines, and their dances were very poorly but not randomly oriented. The bees clearly did not detect the sun's actual location, but approximately half of the bees managed to more-or-less agree on an incorrect dance direction. We have been unable to identify any landmarks that the bees might have used, and we suspect that they may have used the same skyline matching strategy they used in the previous experiment. We cannot analyze the skylines in this case, however, because the skyline at the test site would need to be viewed from above the canopy, 25–30 m above the ground. The bees' first orientation flights at this site did indeed spiral approximately straight upward to the canopy, suggesting that the bees may have viewed the skyline from there. In summary, there is no evidence here for any novel orientation mechanism; the weak orientation we observed in this experiment can almost certainly be explained by Dyer's landscape-based system or, possibly, by the presence of weak, ambiguous celestial cues.

A third experiment, on 22 August 2009, involved another pair of dissimilar sites. The hive and natal site were the same as those in Experiment 2 (Fig. 6A), but the test site was at the edge of a meadow in a valley (Fig. 6B, white dot labeled C). The hive was closed at the natal site at 06:00 h, transplanted, and opened at the test site at 06:25 h. The feeder was moved eastward (white arrow from C in Fig. 6B) and reached 50 m from the hive at 07:10 h. We recorded dances under overcast skies for the next 2 h as the feeder was gradually moved to its final distance of 150 m.

The correct dance direction at the test site was 95 deg (white arrow from C in Fig. 6B, also the thin black line in Fig. 6C), but the dances pointed, on average, toward 169.0 deg (based on 220 dances by N=37 bees, r=0.76; thick arrow in Fig. 6C). Again the bees failed to locate the sun correctly, and again they nonetheless managed to dance non-randomly (P<0.001, Rayleigh test), despite the differences between the natal and test landscapes. In this case, however, the bees' strategy is not difficult to discern: they matched a slightly curved treeline along their flight route at the test site (Fig. 6B, white arrow from C) with a taller, straighter treeline at their natal site (Fig. 6A). The mean dance direction corresponds closely with the treeline-matching prediction (165 deg, thin red line in Fig. 6C), although the scatter is much greater than typically occurs in transplantation experiments involving more similar treelines (Towne and Moscrip, 2008). Once again, our observations are explained by Dyer's landscape-based backup system for locating the sun under clouds.

A final experiment, on 20 July 2012, confirms our interpretation of the previous one. In this experiment, bees were transplanted under overcast skies between the same sites used in Experiment 3 except that the hive was placed 25 m farther from the treeline at the test site (Fig. 6B, white dot labeled D). This 25 m difference caused the bees no longer to recognize the treeline at the test site as a match for the one at their natal site, and the dances, overall, were randomly oriented (Fig. 6D).

Specifically, the hive was opened at 06:20 h under overcast skies at the test site, and the feeder was moved roughly eastward (87 deg, arrow labeled D in Fig. 6B). When the feeder reached 50 m from the hive (at 08:57 h), we began recording dances while continuing to move the feeder to a final distance of 100 m. We recorded dances over the next 4 h under overcast skies with intermittent rain. The two groups of bees were marked either at the bees' natal site 1–10 days before the experiment or at the test site on the day of the experiment, the latter to increase the sample size. There was no reason to expect these two groups to be different, as long as the newly marked bees were experienced foragers, which they almost certainly were. The bees marked at the natal site were poorly but significantly oriented (P=0.003, Rayleigh test; mean direction=319.9 deg, r=0.62, N=14 bees; Fig. 6D, black bars) in a direction that does not correspond with the correct direction (87 deg, thin black line in Fig. 6D), the direction predicted by the treeline match from Experiment 3 (165 deg, thin red line in Fig. 6C) or the best skyline match between the natal and test sites (213 deg, thin red line in Fig. 6D). In this case, the best skyline match was not strong, as can be seen in the panoramic photographs of the sites in the ‘best match’ arrangement in Fig. 7. The other six bees, marked on the day of the experiment, were not significantly oriented (P=0.49, Rayleigh test; mean direction=133.0 deg, r=0.36, N=6 bees; Fig. 6D, gray bars), nor were all 20 bees taken together (P=0.11, Rayleigh test; direction=320.9 deg, r=0.331, N=20; mean vector in Fig. 6D).

Even though the dances overall on this day were poorly oriented, individual bees and individual dances generally were not. Most bees that danced more than once (15 bees) indicated roughly the same geographical direction throughout the observation period (13 bees); only two bees indicated different directions at different times (that is, with a range exceeding 90 deg). In addition, most individual dances were reasonably well oriented, although we did observe more instances of missed wagging runs (where the bee ran two circles between runs) and overshooting than one normally sees. Further, five different bees gave dances that we scored as ‘sloppy’ (six dances), disoriented (one dance) or round (one dance). However, these same five bees gave many (49) well-oriented dances as well. In summary, individual bees on this day appear to have estimated the sun's location as best they could and danced accordingly. Because few of the bees matched the predictions based on the landmarks and skylines, the bees probably used weak or ambiguous celestial cues, although they could have used landscape cues in ways that we do not understand. As in the previous experiments, there is no evidence at all that the bees were able to locate the sun correctly.

When honeybees are transplanted under overcast skies from their home landscape to a visually similar test landscape, the bees attempt to locate the sun using a memory of the sun's compass bearing over time in relation to features of the landscape that seem familiar, which leads the bees to incorrectly locate the sun if the test landscape's features are oriented differently from those at the home site (Dyer and Gould, 1981; Dyer, 1987; Towne and Moscrip, 2008). Here we transplanted bees between landscapes that, we thought, were sufficiently dissimilar to thwart the bees' landscape-based backup system for locating the sun, thereby forcing the bees to use a further backup system, if they had one. In none of our six trials, involving three different sets of landscapes, did the bees locate the sun correctly at our test sites under clouds. However, the bees used their landscape-based system more flexibly than we expected, matching somewhat dissimilar landmarks and skylines. That is, when no other options were available, the bees were sometimes willing to make the best of poor matches between sites. When the landscape matches were poor, however, the bees oriented poorly. We found no evidence that bees have any further, previously unknown backup systems for locating the sun under clouds, such as a memory of the sun's compass bearing over time in relation to the magnetic field.

Our results fully confirm the preliminary findings of von Frisch (von Frisch, 1967) and Dyer (Dyer, 1984), each of whom reported one trial of a similar experiment. When von Frisch transplanted bees into a novel landscape under cloudy skies, none of the bees oriented their dances according to the sun's actual location at the test site. Instead, the dances were disoriented or sloppily oriented in incorrect directions (von Frisch, 1967). These results resemble those of our Experiments 2 and 4 (Fig. 5 and Fig. 6D, respectively), where the bees largely failed to identify similarities between their home and the test landscapes and oriented their dances sloppily. Von Frisch attributed the bees' weak orientation in his experiment to the Lampeneffekt, wherein bees take a lamp or another bright patch in the celestial hemisphere to be the sun (von Frisch, 1967). However, the Lampeneffekt – and other effects of skylight color and brightness on the bees' sun-compass orientation (Edrich et al., 1979; Brines and Gould, 1979; Rossel and Wehner, 1984) – have been demonstrated only in bees dancing on horizontal surfaces, not in free-flying bees attempting to locate the sun under cloudy skies. Von Frisch did not know about the bees' landscape-based backup system for locating the sun, so the Lampeneffekt seemed at the time to be the only reasonable explanation for his bees' (weak) orientation. Now it seems almost certain, however, that von Frisch's bees were attempting to use their landscape-based system to locate the sun, as they did in our experiments.

Dyer's (Dyer, 1984) single trial of an experiment like ours gave results similar to those in most of our trials: the bees did not locate the sun correctly, but the dances were non-randomly oriented in incorrect directions. In Dyer's case, the dances pointed mainly in two different directions roughly opposite each other, both incorrect given the sun's actual location. Dyer speculated that his bees were applying as-yet-unknown rules for matching the home and test landscapes. In our Experiment 3 (22 August, 2009) (Fig. 6C), the bees matched two somewhat dissimilar treelines, a result Dyer might have predicted based on his treeline experiments (Dyer and Gould, 1981; Dyer, 1987). In our Experiment 1 (Figs 1, 2, 3 and 4), by contrast, the bees seem to have matched two somewhat dissimilar skylines, which fulfills Dyer's prediction of a then-unknown rule for matching landscapes. While this skyline-matching mechanism extends our understanding of Dyer's landscape-based system – and there is certainly much more to learn about how ants and bees use landmarks and the skyline in their orientation (Wystrach and Graham, 2012; Zeil, 2012; Palikij et al., 2012; Menzel et al., 2005; Menzel et al., 2012; Collett et al., 2007; Cruse and Wehner, 2011; Cheng, 2012) – it does not represent an entirely novel backup system in bees. Bees evidently cannot determine the sun's location in the absence of useful celestial cues and familiar landscape features. That is, we now seem to know the full depth of the bees' backup sun-compass systems.

The bees' celestial compass systems appear to function hierarchically. In ants, and probably also in bees, the compasses based on direct views of the sun and skylight polarization patterns are separate, and the polarization compass takes precedence (Wehner and Müller, 2006). Further, the sun and skylight compasses take precedence over the landscape-based system: only under overcast skies do the bees use the latter (Dyer, 1987). One exception to this, however, is when bees have been orienting for a time by memory of the solar ephemeris in relation to the landscape and are then exposed to fresh celestial cues, as when the sky clears. Such bees sometimes perform bimodal dances, expressing both compass systems on alternate wagging runs within a single bout of dancing (Dyer, 1987; Towne et al., 2005; Towne, 2008; Towne and Moscrip, 2008). Alternatively, such bees have also been observed to give unimodal dances oriented between the directions predicted by the two compass systems. The dance directions then slowly change over time toward the direction predicted by fresh celestial cues (Dyer, 1987; Towne and Kirchner, 1998). And if the fresh celestial cues persist or become strong, the bees eventually switch to using them alone.

The results of our Experiment 1 (Figs 1, 2, 3 and 4) represent the second instance in which bees appear to have located the sun under cloudy skies by reference to a memory of the sun's course in relation to the skyline (Towne and Moscrip, 2008). In both instances, the individual landmarks around the hives at the test sites were quite different from those at the bees' home sites, and the bees danced according to the best matches between the skyline profiles at the home and test sites. Conversely, in our Experiment 4 neither the landmarks nor the skylines at the natal and test sites matched well, and the bees oriented poorly. We know too little about how bees process skyline information to have predicted the bees' dance orientation in these experiments in advance, however, and more convincing experiments will require such predictions. These transplantation experiments are difficult because they require overcast skies, abundant dancing and special landscapes. In addition, the bees' flights, and therefore their experience before and during the tests, cannot be controlled. An alternative approach to studying this orientation might exploit the bees' ability to orient horizontal dances by reference to familiar landmarks (Capaldi and Dyer, 1995), substituting the kind of artificial skylines already used with ants (Graham and Cheng, 2009a) for the landmarks that have been used with dancing bees (Capaldi and Dyer, 1995). Yet another approach might use measurements of the bees' departure bearings from feeders, a method used recently by Najera et al. (Najera and Jander, 2012; Najera et al., 2012) in studies of the bees' spatial orientation.

Use of the skyline in insect orientation was first suggested by observations of ants (Fukushi, 2001) and bees (Southwick and Buchmann, 1995; Towne and Moscrip, 2008) and definitively demonstrated recently in ants (Graham and Cheng, 2009a). The first explicit demonstrations that bees use the visual panoramic context in orientation (Collett and Kelber, 1988; Collett and Zeil, 1997; Collett et al., 2002), combined with Möller's suggestion (Möller, 2002) that the ground/sky contrast is well suited for use in orientation, precipitated a series of studies on the significance of the visual panorama (Collett et al., 2003; Collett et al., 2006; Stürzl et al., 2008; Wystrach et al., 2011a; Wystrach et al., 2011b; Zeil et al., 2003), and the skyline in particular, in the view-matching orientation strategies of bees (Pahl et al., 2011) and ants (Baddeley et al., 2011; Baddeley et al., 2012; Basten and Mallot, 2010; Graham and Cheng, 2009a; Graham and Cheng, 2009b; Reid et al., 2011; Philippides et al., 2011; Wystrach and Graham, 2012). The skyline-matching mechanism appears to be especially robust and flexible, as the skyline is distant and easily detectable (Möller, 2002), and bees evidently do not rely on the skyline's details (Fig. 4), so that any natural changes in the skyline would be unlikely to affect them. Several excellent reviews have been published on these subjects recently (Cheng, 2012; Collett et al., 2007; Wystrach and Graham, 2012; Zeil et al., 2003; Zeil, 2012).

The bees' lack of a further backup system does not compromise the effectiveness of the dance communication under cloudy skies. First, the skyline-matching mechanism appears to be especially robust and flexible, as discussed above. Second, most or all dancers and potential recruits in a colony appear to know the sun's position over time in relation to the entire landscape panorama near the nest, even routes with which the bees have had no specific experience (Towne and Moscrip, 2008). Thus it is probably rare indeed that bees cannot dance and interpret dances correctly using the sun compass in their natal landscape, even under the most difficult conditions. This, then, could explain functionally why bees either do not have or do not use a memory of the sun's course in relation to the magnetic field as a further backup system for cloudy days: they do not need it. It remains possible, even likely, that bees use their magnetic compass when they first learn the landmarks (Collett and Baron, 1994; Frier et al., 1996) and skyline panorama (Towne and Moscrip, 2008) around their nests. But once these and the sun's pattern of movement in relation to them are learned, the celestial compass evidently takes over as the bees' primary compass (Dickinson, 1994; Frier et al., 1996).

Our first experiment here (Fig. 2) provides a possible alternative explanation for Towne's (Towne, 2008) results in which bees were transplanted from their natal landscape to a dissimilar one, and the bees appeared to re-learn the sun's course in relation to the novel site over the next few days. This re-learning was measured by testing the bees under overcast skies at a third site that was a rotated twin of the second. However, the natal and test landscapes Towne used were the same two landscapes we used in our first experiment here (Fig. 2A,B), and our bees danced in approximately the same direction as Towne's bees, even though they had no opportunity to re-learn the sun's course. Instead, our bees evidently used the skyline-matching mechanism discussed above, which could explain Towne's (Towne, 2008) results without the re-learning he inferred. However, we have now repeated Towne's experiment using different sites and control bees transplanted directly between the natal and test sites, and we have confirmed Towne's original conclusion: bees transplanted between dissimilar sites re-learn the sun's course in relation to the novel site (J.R.K. and W.F.T., submitted). Bees do not re-learn the relationship between the sun's course and the landscape, by contrast, when they are transplanted between strongly similar sites (Towne and Kirchner, 1998; Towne et al., 2005).

We thank Paul M. Bauscher, Greta Campbell, and Walter and Doris Fink for allowing us to work on their properties; Lauren Quintrell and Sam Towne for assistance in the field (2009); and two anonymous reviewers for many suggestions that greatly improved the manuscript.