In the present article, feature selection/weighting based on SVM was employed to improve the algorithm of choosing reference spectrum through a multi-objective optimization approach proposed in reference. Based on the sensitive analysis, half of features having low weights in SVM classification model were eliminated iteratively. Two criteria, matching accuracy and classification confidence, were used to select the best-performing feature subset. Three scenarios were designed: (1) only feature subset selected by SVM was used; (2) both feature subset and global weights were used, in which global weights were the coefficients of selected features in the SVM classification model; (3) both feature subset and local weights, which changed with the distance of a sample point to the SVM separation plan, were used. Experiment executed on the popular Indiana AVIRIS data set indicate that under all the three scenarios, spectral matching accuracies were increased by 13%-17% compared to the situation without feature selection. The result obtained under scenario 3 is the most accurate and the most stable, which can be primarily attributed to the ability of local weights to accurately describe local distribution of spectra from the same class in feature space. Moreover, scenario 3 can be regarded as the extension of scenario 2 because when spectra far away from the separation plane are selected as reference spectrums for matching, the features' weights will not be considered. The results obtained under scenario 1 and 2 are very similar, indicating that considering global weights is not necessary. The research presented in this paper advanced the spectrum analysis using SVM to a higher level.