In order to quickly and accurately detect the phosphorus content information of rice leaves in the cold northern region, analyze the growth of rice, and provide the basis for precision fertilization and scientific management of rice fields, this study takes rice in the cold northern region as the research object, based on the plot experiment, and uses the marine optical HR 2000+ spectrometer to obtain the hyperspectral reflectance data of rice leaves, the content of phosphorus in rice leaves was determined by vanadium molybdenum yellow colorimetry.SG smoothing and multiple scattering corrections (MSC) were used to preprocess rice leaf hyperspectral data. The spectral data were selected by using SPA and UVE. Eleven features were screened by the SPA algorithm, including 6 in the visible band (411, 420, 428, 442, 467 and 689 nm) and 5 in near-infrared band (797, 850, 866, 965 and 976 nm). A total of 47 features were obtained by the UVE algorithm, which was all located in the visible band range and distributed between 405 and 603 nm. Taking the characteristic reflectance selected by the two methods as input, three inversion models of phosphorus content in rice leaves were constructed and analyzed, including extreme learning machine (ELM), BP neural network and wolf pack algorithm (WPA-BP).The results show that the verification set R2 of the three models constructed with the characteristic reflectivity filtered by the UVE algorithm as the input is between 0.705 2~0.724 5 and RMSE is between 0.017 4~0.020 4; Under the condition of the same inversion model, the prediction effect of the model constructed by using the characteristic reflectivity filtered by SPA algorithm as the input is good. The verification set R2 of the three models is between 0.726 4~0.829 3, and RMSE is between 0.018 0~0.021 1; In addition, when using the features screened by the two algorithms for modeling, comparing the prediction results of the three models, it is found that the accuracy of the BP neural network model optimized by the wolf swarm algorithm is significantly higher than that of the ELM and BP neural network, and the determination coefficient R2 of the verification set is 0.803 4 and RMSE is 0.018 0. Because of this, the combination of SPA and WPA-BP has certain advantages in a hyperspectral inversion of phosphorus content in rice leaves in the cold northern region, which can be used as a reference for the detection of phosphorus content in rice leaves and accurate quantitative fertilization.