Abstract: Abstract: Traditional methods of obtaining phosphorus content of citrus leaves are time-consuming procedures with complex operations which can be harmful to citrus trees. More over, traditional methods can not meet the demand of rapid and non-destructive monitoring of phosphorus content in large-scale citrus orchards. In this paper, we presented several models suitable for phosphorus content prediction in 4 growth periods using hyperspectral information. The experiments were conducted in the Crab Village of Luogang District, Guangzhou City, Guangdong Province, and the samples were 195 citrus trees planted. During 4 growth periods, i.e. germination, stability, bloom and picking period, hyperspectral reflectance of citrus leaves was respectively measured by spectrometer (ASD FieldSpec 3), and at the same time, phosphorus content of citrus leaves was obtained by using traditional chemical method. Owing to the high dimensionality and redundancy of raw data, an enhanced algorithm was provided based on manifold learning to deal with the high-dimensional spectral vectors for dimension reduction and feature extraction. First of all, the parameters of wavelet de-noising, which was applied to reduce the high-frequency noise, was determined through orthogonal test, and then 5 manifold learning algorithms, i.e. laplacian eigenmaps (LE), locally-linear embedding (LLE), local tangent space alignment (LTSA), isometric mapping (Isomap) and maximum variance unfolding (MVU) were applied to reduce dimension and extract features for de-noising spectrum. The 5 corresponding prediction models of support vector regression (SVR) for phosphorus content of citrus leaves were established based on their features. Besides, we compared the modeling results of different spectral forms. Some critical conclusions were obtained. First, the optimized parameter combination of wavelet de-noising through orthogonal test was: "coif2" as wavelet basis function, the number of decomposition layer being 7 and "heursure" as the threshold, respectively. Second, the experimental results revealed that these 5 manifold learning algorithms were effective for phosphorus content estimation of citrus leaves. When the raw spectrum was used as the input vector, the Isomap-SVR model achieved better performance than other models; the coefficients of determination for the calibration set were 0.9383, 0.9614, 0.9611, 0.9516 and 0.9430, and the corresponding values of root mean square error (RMSE) were 0.0548, 0.0503, 0.0456, 0.0534 and 0.527 at germination, stability, bloom, picking period and whole growth period, respectively; for the validation set, the coefficients of determination were 0.8866, 0.8923, 0.9236, 0.9005 and 0.8870, and the values of RMSE were 0.0710, 0.0688, 0.0583, 0.0667 and 0.0704, respectively, which meant high usability for the industry. Third, when first derivative spectrum was used as the input vector of the samples with wavelet de-noising, in our research, the Isomap-SVR model achieved the best result and the coefficients of determination for calibration set were 0.9383, 0.9614, 0.9611, 0.9516 and 0.9430 respectively, and the corresponding values of RMSE were 0.0518, 0.0405, 0.0408, 0.0458 and 0.0499 respectively at germination, stability, bloom, picking period and whole growth period; and for the validation set, the coefficients of determination were 0.8913, 0.9107, 0.9373, 0.9135 and 0.8949, and the corresponding values of RMSE were 0.0703, 0.0645, 0.0522, 0.0634 and 0.0659 respectively. Finally, our research proves the feasibility of monitoring phosphorus content of citrus leaves, and may provide a theoretical basis for growth monitoring and nutritional diagnosis of citrus trees.