The profit of low-cost, multispectral imaging systems in estimating spectral power distributions has been widely studied. There are various mathematical methods available (PCA, Wiener's estimation method, spline interpolation, MDST, among others) which permit the accurate reconstruction of a spectrum from the response of a small set of sensors. One important issue in this task is the influence of noise, its propagation through mathematical transformations and how the selection of the sensors of the multispectral system, combined with the spectral estimation algorithm chosen, may reduce its influence. We report here on four different spectral recovery methods that reconstruct skylight spectra from the responses of three Gaussian sensors (the spectral profile of which is a Gaussian curve). The sensors are searched for using a simulated annealing algorithm, and they are optimized so that they give the best possible spectral and colorimetric reconstructions, even in the presence of noise. We show here how the accuracy of the reconstructions is influenced by the recovery method chosen.