In previous works [1,2] authors developed a series of computational simulations in order to study the real behaviour of a multispectral imaging system aimed at skylight spectral recovery. We took into account the effect of several parameters on the spectral shape of the optimum sensors found for recovering skylight spectra. Hence, we studied the influence of different kinds of noise, number and kind of sensors, linear bases, number of representative vectors, size of the training set and the spectral estimation method. In this work we go along these lines by implementing the five optimum sensors found in a realizable case by using a monochrome CCD camera and a Liquid Crystal Tunable Filter (LCTF). We show that the computational simulations resulted in a very realistic study of the behaviour of this practical multispectral system since the spectral recoveries obtained with the real optimum system are quite acceptable for scientific purposes where skylight spectra are used. Hence, we have constructed a real multispectral system for imaging skylight, that could be used later to obtain information about climate parameters like the Angström exponent or the optical depth [3] in every pixel of the image.