Abstract: Abstract: To classify muskmelons quickly and accurately based on dielectric spectroscopy, dielectric properties of 4 kinds of melons (a total of 246) were measured with network analyzer over the frequency range from 20 to 4 500 MHz. The samples were selected from 4 different greenhouses in Yangling, Shaanxi Province, which had similar shape and size, and had no injury and disease. All samples were divided into calibration set and validation set with a ratio of about 3:1 based on Kennard-Stone method. Methods of support vector machine (SVM) and extreme learning machine (ELM) were applied to establish discriminative models of muskmelons. We chose 2 different variable selecting methods as pre-processing methods before modeling. One method was principal component analysis (PCA) for data dimension reduction, and the other was successive projections algorithm (SPA) for characteristic variables selecting. The model validating effects after the processing of PCA and SPA were used to compare with that with no pre-processing; besides, directly modeling with full frequencies (FF) spectra data was also adopted. The results were shown as below: 1) All discriminative models under FF, PCA and SPA methods could be used for classifying muskmelons. The total correct rate of each validation set reached over 96%, and the ELM modeling method was better than SVM method as a whole. 2) The models based on the FF method retained all original information of the frequency spectra data, so it had the highest validation correct rate, up to 100%. But its stability and reliability were not good enough because of the existing interference information. Under the PCA method, the accumulating contribution rate of the former 10 principal components extracted from all variables approached to 99.99%, which well reflected original information while simplifying the model in some degree and improved performance of models, however, the results were not very stable and the total correct rate of 2 models was much lower than others, up to 96.72% and 98.36% respectively. Seventeen characteristic variables were selected by the SPA from all 202 variables for modeling, which not only simplified the model and improved its performance, but also had the higher accuracy. Therefore, the SPA method was more suitable for the variables selecting based on dielectric spectrum. 3) In all models, SPA-ELM had the minimum misjudgments and the highest total correct rate, which was more suitable for classifying muskmelons according to dielectric frequency spectra. Therefore, it's feasible to classify muskmelons based on dielectric spectrum by the modeling methods of SVM and ELM. It also shows that the dielectric spectrum technology can be used to do more research on muskmelon classification and grading, and provides the new theory and methods for future research about nondestructive detection of muskmelons.