Multispectral images of natural scenes have been used to obtain a recovery matrix that transforms directly R, G, B values to spectral radiance without using linear models for spectral reflectances and illuminants. The R, G, B are simulated digital counts from a standard CCD camera. This recovery method had not been tested in natural scenes with uncontrolled illumination and for spectral radiance. We have used three different test data sets to check the method's accuracy, and three different quality measures to test the similarity between recovered and original spectral radiances (Goodnes-of-Fit, Root-Mean-Square-Error and ΔE*_ab colour difference). With this simple colorimetric system it was found that natural spectra could be recovered with a quality that is adequate for many applications.