Reflectance estimation from RGB data in natural scenes is studied computationally including the use of different unsupervised classification techniques to divide the RGB data into a number of subgroups with similar characteristics to test if these techniques lead to any improvements in the quality of the spectral signals obtained. The direct pseudoinverse method for recovery of spectral signals from RGB values is used for each subgroup and the similarity of the recovered spectral data to the original sets is tested by different quality indexes. Weighted mean results according to the number of components of each subgroup are compared with mean results obtained for the whole RGB data set (with no classification algorithms used as preprocessing step). Different algorithms and number of classes are tested for noise-free and noisy data. In addition, the use of an color filter in front of the camera lens is introduced in the computations to study spectral recovery from six instead of three RGB values for each spectral reflectance. The best results are obtained for 8 classes and a probabilistic approach clustering algorithm. Quality decreases when a high level of noise is added to the data, and the use of a color filter only helps to improve results for noise-free data.