Multimodal skylight information improves the estimation of the celestial compass: insights from a hardware implementation

Abstract

The most common problem in navigation is the accumulation of directional errors. Insects solve this problem using a neural compass based on external celestial cues. In previous work, we hypothesised a computational model of a neural circuit that converted polarised skylight to compass direction. The model was constrained by the insect’s sensor array and operated under a simulated sky. Using only the polarisation pattern of the sky, it could accurately estimate the solar azimuth, elevation, and the confidence of its estimations.

In order to estimate the performance of this model in real-sky conditions, we built a compass sensor prototype with eight polarisation sensitive units arranged on a ring (simplifying the dome structure of the original model) and elevated by 45 degree. Following the model’s description, each unit integrated the input of two photodiodes to separate the light intensity from the degree of polarization. The photodiodes were sensitive to ultraviolet (UV) light and placed under linear polarisation filters, which were aligned with the units’ meridian and oriented perpendicularly to each other. We collected the responses of the photodiodes while rotating the compass sensor under different sky conditions: clear or overcast sky, under trees or solid canopies, and at different times of the day. The compass model used these to estimate the heading direction with respect to the sun. Finally, we estimated the heading direction using different modalities of light (intensity or degree of polarisation) and compared their performance.

Light intensity seemed particularly useful as a cue under clear sky conditions and when the sun was clearly visible by the sensor. The degree of polarisation made the estimations of the compass more accurate when the sun was hidden, under thin clouds or tree covers. Simple retinotopic integration of the two modalities outperformed both the intensity- and polarisation-only models and utilised the advantages of both. Our results suggested that the compass design and computational model could successfully predict the location of the sun in most solar elevations, cloud covers, occlusions and atmospheric conditions. Also, this hardware implementation allowed for interesting predictions on the different stages of the polarisation pathway in the insect brain.