Soil nutrients are the main attributes of the soil, affecting the quality of the soil and the growth of plants. This study aims to use hyperspectral remote sensing data to construct an inversion model of soil organic carbon, total nitrogen, total phosphorus, and total potassium content in the source area of the Three Rivers and draws some conclusions and laws suitable for soil nutrients in the source area of the Three Rivers. It uses a variety of mathematical transformations to optimize the soil spectral data. It eliminates or reduces the influence of irrelevant noise, and improve the sensitivity of soil nutrients to the spectrum. It uses multiple linear regression (MLR), partial least square regression (PLS), support vector machine (SVM), and random forest (RF) to establish a hyperspectral inversion model of soil nutrients. Comparing the simulation results of the four models, it is found that the overall accuracy of RF is the highest. Among the four soil nutrient indexes, the multivariate scattering correction-first-order differential-random forest (MSCFD-RF) model has a higher prediction accuracy for soil organic carbon and total nitrogen, and the multivariate scattering correction-random forest (MSC-RF) model is more accurate The simulation results of potassium are better, and the R2 of the verification set accuracy are 0.8739, 0.8592, and 0.8865 respectively, which shows that the prediction effect is relatively ideal. The verification set accuracy R2 of all inversion models of soil total phosphorus is lower than 0.2, and the RPD value is lower than 1.4. It is impossible to predict the total phosphorus content, and the model fitting effect is not ideal.