The mountain meadow on the north slope of the Tianshan Mountains has the highest grassland productivity, and the grassland degradation is serious. The classification and recognition of grassland vegetation play an important role in monitoring the background status of the grassland ecosystem. It is also the key to carrying out ecological reconstruction, which can quickly, accurately, and effectively evaluate grassland the dynamics and degree of grassland degradation. In this paper, we explored the classification method in grassland vegetation of the typical mountain meadow vegetation in the middle section of the north slope of Tianshan Mountain in Xinjiang. Firstly, a hyperspectral imaging spectrometer obtained original reflectance spectra of typical vegetation in four key growth periods (SOC710VP). Then, Savitzky-Golay filtering and the minimum noise fraction transformation (MNF) were used to smooth and reduce the dimensions of the spectrum data. Thirdly, classification models were established by the support vector machine (SVM), the backpropagation artificial neural network (BP-ANN) and the spectral angle mapper (SAM). Finally, a comparative analysis of the classification results from three models was made. The results showed that the dimension reduction and noise removal of grassland vegetation hyperspectral data could be effectively carried out by using the S-G filter and MNF transform preprocessing method. This processing reduced the redundancy of data and shortened the classification time while obtaining a smoother classification image. The parameters such as “green peak”, “red valley”, and “red edge” of mountain meadow vegetation varied greatly in different seasons. The spectral curve characteristics in the vigorous vegetation growth period (from April to May) were easier to distinguish than those in the withering date. Thus, the classification accuracy was higher in this period. The overall classification accuracy of the SVM model exceeded 90%, and the Kappa coefficient exceeded 0.9 in the green-up date (April) and tillering stage (May). Based on the SVM model, the classification accuracy of the polynomial kernel function was higher in the vigorous growth period (from April to May), and the radial basis function (RBF) showed better performance in the mature period (from June to September). BP-ANN had higher classification accuracy in the tillering stage, the overall classification accuracy was 91.07%, and the kappa coefficient was 0.89. However, the classification effect was general in other periods. Moreover, the classification time was still longer than that of SVM although after the reduction of MNF transformation dimensionality. SAM had the fastest classification speed, but the classification accuracy was low in each growth stage. The highest value was 77.80% of the overall classification accuracy in tillering stage, and the kappa coefficient was 0.73. Therefore, the SVM classification model using the polynomial kernel function was suitable for classifying and recognising mountain meadow vegetation, which had complete classification category results, higher accuracy and relatively few misclassification. It was a better classification method than BP-ANN and SAM.