Traditional methods of obtaining nitrogen content of citrus leaves are time-consuming, and the process is cumbersome and harmful to citrus leaves, which need proficient experiment techniques and amounts of instruments, equipment and chemical reagents. According to the high dimensionality and redundancy of origin spectral reflectance, a nitrogen content obtaining method of citrus leaves was provided based on manifold learning algorithm which was applied to the high-dimensional spectral vectors for dimension reduction and feature extraction. During four different growth stages, corresponding to germination, stability, bloom and picking stages, spectral reflectance of citrus leaves were measured by the ASD FieldSpec 3 spectrometer, respectively, and at the same time, nitrogen content of citrus leaves was obtained by using Kjeldahl method. For data processing, firstly the parameter combination of wavelet denoising which was used to the high-frequency noise removal was optimized through orthogonal test, and then the principal component analysis (PCA), multidimensional scaling (MDS), locally-linear embedding (LLE), isometric mapping (Isomap) and laplacian eigenmaps (LE) manifold learning algorithms were applied to extract features of original spectrum and denoised spectrum. Finally, the five corresponding support vector regression (SVR) prediction models of nitrogen content for citrus leaves were established based on their features. Experiment results reveal that the five manifold learning algorithms can be effectively used to predict nitrogen content of citrus leaves, which provides theoretical basis for obtaining nitrogen content of citrus leaves rapidly and non-destructively, as well as in growth monitoring and variable-rate fertilization.