Abstract: In order to quickly test the soil nutrient contents (SOM, TN, TP and TS), the authors collect 117 soil samples at 0-20 cm depth from Chuangye Farm in Jiansanjiang reclamation area as research objects. First derivative (FD), logarithmic reciprocal (RL), first derivative of reciprocal (FDR), multivariate scattering correction (MSC) and continuum removal (CR) transformations are performed on the raw spectral reflectance (R). By analyzing the correlation between the six spectral variables and soil nutrient content, the bands that are significantly correlated at the α=0.01 level are adopted as characteristic bands, and the methods of stepwise multiple linear regression (SMLR), partial least squares regression (PLSR) and back propagation neural network (BPNN) are used respectively to establish hyperspectral prediction model of SOM, TN, TP and TS. The model is evaluated by R², RMSE and RPD. The results show that the soil nutrient content prediction models established by PLSR and BPNN are superior to that by SMLR. The PLSR and BPNN methods can well predict the organic matter and total nitrogen content, and roughly estimate the total sulfur content. Only the CR-BPNN method can roughly estimate the total phosphorus content. The models with the best prediction effect on SOM, TN, TP and TS are, respectively, MSC-PLSR, MSC-PLSR, CR-BPNN and FDR-BPNN, with the validation set determination coefficients of 0.86, 0.75, 0.56 and 0.67 respectively.