In order to investigate an effective method for rapid discrimination of plum fruits at different maturity levels, plum fruits were used as the research object of this study, and spectral information (420~1000 nm) of plum fruits at different maturity levels (unripe, mid-ripe, ripe, over-ripe) was collected using hyperspectral imaging technology. A combination of smoothing and standard normalized variate (SNV) was applied to pretreat spectral data. Then, the partial least squares method(PLS) model was established for comparing the accuracy of different discriminant models, through using the pre-treated full spectrum (FS) data, the principal components extracted by the principal component analysis (PCA) method, and the feature wavelengths extracted by the successive projection algorithm (SPA) technique as input variables. The results revealed that the model established by FS-PLS had the highest discriminative accuracy and the accuracy rate reached 91.88%. However, considering the amount and complexity of the experimental calculations, the accuracy rate of the model established by SPA-PLS was the best with the comprehensive accuracy reaching 91.25%. This study provides a new theoretical basis for discriminative detection of plums maturity.