A comparison of photographic and spectrometric methods to quantify the colours seen by animal eyes

Abstract

Abstract Quantifying how colours stimulate animal eyes is an essential first step in many areas of biology, from behavioural and visual ecology to neuroethology. Although multiple methods exist, their relative accuracies remain untested, leaving researchers unsure which methods are most accurate and which might be unacceptably inaccurate. Here, I compare measurements from reflectance spectrometry and four camera‐based techniques to ground‐truth spectroradiometric measurements of radiance, using eight diverse visual systems. Because errors may compound through successive calculations, I calculate not only relative quantum catches but also chroma and colour contrasts. I find that filters mimicking photoreceptor spectral sensitivity curves, through either custom fabrication or weighted additive combinations of bandpass filters (‘computational filters’), as well as reflectance spectrometry, are the most accurate techniques, with most R2 > 0.98. Statistical filters trained on the library of reflectance spectra in the multispectral image calibration and analysis (MICA) toolbox lag behind in accuracy, particularly for dark colours, with R2 values dipping as low as 0.23. However, the performance of these filters is vastly improved (all R2 > 0.96) if one (a) uses the full UV–VIS range of camera and animal vision, even for animals lacking UV photoreceptors and (b) expands the library of training spectra to include simulated low‐reflectance spectra. Statistical filters trained on either a 24‐ or 45‐colour chart (an alternative method offered by the MICA toolbox) vary wildly in accuracy, with R2 values ranging from 0.11 to 0.94, and should not be trusted. For those requiring point measurements only, reflectance spectrometry is highly accurate within the constraints described in the main text. Those using photography and desiring the highest levels of accuracy should opt for custom‐fabricated or computational filters. Statistical filters trained on the MICA toolbox's library of natural spectra can be expected to be reasonably accurate for many combinations of visual systems and stimuli, but only if (a) the full UV–VIS spectrum is used and (b) the training library is expanded to include simulated low‐reflectance spectra. Statistical filters trained on a colour chart are highly unreliable and their use should be discontinued.