The accurate acquisition and prediction of physiological information can provide a basis for the fine management of planting. The traditional soybean physiological information inversion methods have low detection efficiency, cumbersome operation process and mostly damage detection. To this end, this paper uses hyperspectral technology to establish a rapid non-destructive inversion method for soybean physiological information. The leaves of soybean flowering and pod-forming period were taken as research objects, and the hyperspectral, chlorophyll content, net photosynthetic rate and photosynthetically active radiation data were obtained on 2 dates (D1 and D2). First, multiple scatter correction (MSC), standard normal variable transformation (SNV), first derivative (FD), second derivative (SD), Savitzky-Golay smoothing (SG), MSC-SG-FD, MSC-SG-SD, SNV-SG-FD and SNV-SG-SD methods are used to preprocess the original spectral data, then establish a full-band model by partial least squares (PLS), compare and analyze, and select the optimal preprocessing method. The feature wavelengths are filtered and extracted by competitive adaptive reweighted sampling (CARS), successive projections algorithm (SPA) and correlation coefficient (CC), respectively. Finally, the preferred preprocessing method and the characteristic wavelength variable are modeled and compared by PLS. The correlation coefficients Rc and Rp of the correction set and the prediction set are used as model evaluation indexes, and the inversion model with the highest correlation with soybean physiological information is finally selected. The results showed that the R_c and R_p of the full-band PLS model with chlorophyll content established by MSC-SG-FD pretreatment were the highest, 0.909 and 0.882 (D1), 0.909 and 0.880 (D2), respectively, the R_c and R_p of the full-band PLS model with light energy utilization established by SNV-SG-FD pretreatment are the highest, 0.913 and 0.894, 0.902 and 0.869, respectively, which shows the highest model performance characteristics compared with the original and other pre-processed models. Further comparing the modeling of the three characteristic wavelength extraction methods, it is found that the variables selected by the SPA algorithm can compress the modeling variables of the chlorophyll content inversion model from 512 to 20 (D1) and 23 (D2), and the variable compression rate is as high as 96.09% and 95.51%. At the same time, the modeling variables of the light energy utilization inversion model can be compressed to 27 and 37, and the variable compression rate is as high as 94.73% and 92.77%. Finally, the optimal modeling method for inversion of chlorophyll content is MSC-SG-FD-SPA-PLS with R_c values of 0.944 (D1) and 0.941 (D2), R_p values of 0.911 and 0.903, and the optimal modeling method for inversion of light energy utilization is MSC-SG-FD-SPA-PLS with R_c values of 0.929 (D1) and 0.925 (D2), R_p values of 0.912 and 0.907 The model has high precision and can provide technical support for large-area detection of physiological information.