Visible–short-wavelength infrared reflectance spectral data is the most essential information for mineral discrimination, as required in technical applications for remote sensing, including resource exploration, weathered-zone mapping, and tunnel-face stability assessment. This study aims to specify a reliable method for estimating the content ratios of endmember minerals from the scaling-effect-containing reflectance spectral data of mixed minerals or different chemical compositions. Eleven samples of calcite and montmorillonite mixed in weight ratios from 0 to 1 (pure montmorillonite and calcite, respectively) were prepared in 0.1 increments. Fully, scaled, log and continuum removal (LCR), and normalized constrained least-square unmixing (FCLSU, SCLSU, LCR-CLSU, and N-CLSU, respectively) and the Euclidean distance and absolute value of the principal component obtained using regularization linear discriminant analysis (RLDA) were used for selecting spectral features, and their estimation accuracies were compared. In another trial to improve the estimation accuracy, LCR processing was combined with spectral feature selection to use only the reflectance data at the selected features’ wavelengths. The estimation accuracy was evaluated based on the smallness of the root-mean-square error in the estimated calcite ratios. SCLSU and LCR-CLSU produced small errors in the calcite ratios estimated from scaling-effect-containing reflectance spectral data. The estimation accuracy was further improved (almost doubled) by combining RLDA for selecting wavelengths with LCR processing because RLDA reduces the importance of the reflectance wavelengths absorbed by the montmorillonite’s structural water, the content of which substantially impacts the estimation accuracy. Consequently, the selection of spectral features can probably contribute to the highly accurate estimation of the content ratios of endmember minerals in multimineral mixtures, which are our next target materials.