In order to quickly analyze varieties of tomato via space mutation breeding with near infrared spectra, characteristics of the pattern were analyzed by partial least square. The model was built with radial basis function neural network and regarded the compressed data as the input of neural network input vectors. The model regarded the compressed data as the input of neural network input vectors and the training process was speeded up. For one hundred and five fruit samples of CK, M1 and M2 the training model was built. Forty five samples formed the prediction set. The discrimination rate of these two models achieved 95.6% and 97.8%. It offered a new approach to the fast discrimination of varieties of tomato via space mutation breeding.