Near infrared diffuse reflection spectra of 15 species’ 171 samples of Dendrobium combined with chemometrics statistical analysis were used to build prediction model, in order to discriminate different species of Dendrobium quickly and nondestructively. Hotelling T2 was applied to stability analysis of spectrum of 5 random drawing samples, and the results showed that the samples spectrum possessed good stability. Orthogonal test L24（2×4×3×8） was designed to optimize optical path type, spectral band, derivative and smooth. The result of orthogonal test was analyzed by principal component analysis, which revealed that when 6 500～4 000 cm-1 spectral band was applied, and with multiplicative scatter correction, second derivative, Norris smooth, and the number of principal components 7, the spectrum distinguishing accuracy was 100%. With the optimized condition of orthogonal test as the input value of partial least squares discriminant analysis and random drawing 123 samples as calibration set to establish the prediction model, and the rest 48 samples as prediction set were use to assess the property of the prediction model, the results indicated that the accumulating contribution rate of the first 3 principal components of the model was 99.36%, the identification of the standard deviation was ±0.1, and the correct recognition rate of the model was 97.92%. The results were satisfied. The method provided a new way for the rapid identification of different species of Dendrobium, and also supplied a reference for the authentication of medicinal plants.