In this study, the principal component analysis (PCA) technique and its nonlinear version (NLPCA) are employed for the compression and reconstruction of spectral data. The reflectance spectra of 1269 matt Munsell color chips are used as original dataset in 400 to 700 nm with 10 nm intervals. The hidden patterns of spectral data are determined by employing the classical PCA as well as its nonlinear version. Different numbers of feature vectors are used in both methods and the results compared by using the root mean square error (RMS), the goodness fit coefficients (GFC) as well as the color difference values under D65 illuminant and 1964 standard observer. Results show the priority of NLPCA over the PCA in low-dimensional spaces i.e. up to 4 basic functions, while different results are observed in higher-dimensional spaces.