Discrimination of crop’s nitrogen level can contribute to reasonable and effective fertilization. Lettuces of various nitrogen levels were planted in three soilless nutrient solutions of different nitrogen concentrations. In the rosette stage, 84 lettuce leaves of each nitrogen level were collected and scanned by the hyperspectral imaging acquisition system. In every hyperspectral image of lettuce leaf, four different positions of 60×60pixel were selected as regions of interest (ROI). The average spectral data of the ROI were used as the original spectra of the leaf samples. The original spectra were preprocessed by the standard normal variate correction (SNV), and their dimensionalities were reduced through principal component analysis (PCA). ELM algorithm was used to establish model for the training samples, and then was compared with BP algorithm model and SVM algorithm model. The results show that the running time of ELM model is 0.62304s and its classification accuracy rate is 100%. During the same running time, the classification accuracy rate of ELM model is higher than that of SVM model. At the same classification accuracy rate, the running time of ELM model is shorter than that of BP model.