This paper points out KICA-NFCM algorithm to identify 4 alpine grassland types using HSI hyper-spectral images, by the comparative study of three spectra and two algorithms. Spectral reflectance data for stipa purpurea, kobresia tibetica, little kobresia and kobresia pygmaea was collected from HSI images, based on field investigation and inspection on the spot. Logarithm transformation and derivative transformation were used in the original spectra of 4 alpine grassland types. Sensitivity bands were determined for original spectra data, first-derivative spectra and logarithmic transform spectra, after the application of waveform analysis, one-way ANOV and correlation analysis. Then, sensitivity bands were imported into KICA-NFCM algorithm to identify 4 alpine grassland types mentioned above. For the sake of contrast, ICA-FCM algorithm was tested too. For original spectra data, first-derivative spectra, and logarithmic transform spectra, sensitivity bands were as follows: 788~925, 711~742, 669~682 and 788~925 nm respectively. Based on original spectra data, first-derivative spectra, and logarithmic transform spectra using KICA-NFCM algorithm, overall classification accuracy and KAPPA coefficients were as follows: 75.38%, 0.685; 81.26%, 0.752; 87.65%, 0.823. In contrast, overall classification accuracy and KAPPA coefficients were as follows: 64.39%, 0.569; 67.74%, 0.604; 73.14%, 0.662, based on three types of spectra using ICA-FCM algorithm. Results show that comparing with original spectra data and first-derivative spectra using ICA-FCM algorithm, logarithmic transform spectra using KICA-NFCM algorithm can make a more accurate and efficient identification of 4 alpine grassland types mentioned above, as well as the “salt and pepper noise” was suppressed in classed images. In contrast, ICA-FCM algorithm decreased boundary precision of patch in classed images and region consistency. Using “logarithmic transform spectra / ICA-FCM algorithm” proposed in this paper, the above 4 alpine grassland types in Naqu prefecture can be identified more accuracy. This method provides technical foundations for the development of hyper-spectral imaging observation for alpine grassland.