In this study， the terahertz time-domain spectroscopy （THz-TDS） technology was used to conduct experimental tests on seed samples of astragalus japonica. We obtained the terahertz time-domain spectra of five kinds of Astragalus adsurgens Pall. seeds in the effective frequency range of 0.2-1.2 THz， and used the fast Fourier transform analysis to study the optical parameters such as the absorption coefficient and refractive index of each grass-seed sample. It was found that in the effective frequency range， the peak intensity and delay time of the time-domain spectrum of the samples were different， and the average absorption coefficient and standard deviation of each spectrum line were significantly different. In addition， the average refractive index of the samples was significantly different. At the same time， this study proposes a hybrid model of optimized experimental data that combines principal component analysis （PCA） with random forest machine learning （RF）. Based on the terahertz refractive index spectrum， 200 datasets of five forage species were statistically calculated， and the calculated results were compared with the calculated results of the RF model. The results show that the average classification accuracy of principal component analysis-random forest （PCA-RF） in the mixed model is 91.20%. Compared with the RF model， both total average classification accuracy and the classification accuracy of each sample of the PCA-RF model are better than those of the RF model. The study shows that the PCA-RF model combining THz-TDS with the hybrid machine learning algorithm can be used as an effective method for the nondestructive identification of the authenticity of forage grass seeds. In particular， it can be used for the classification of forage grass varieties of the same family with little difference.