Random sample approaches were employed for the sampling scenario with 41 spots sampled in topsoil (0-20 cm) in Red Soil ecology experimental station,Poyang Lake ecology experimental station,and Qianyanzhou ecology experimental station in Jiangxi Province by global position system (GPS). The hyperspectral reflectance spectra of soil samples were measured in ultraviolet,visible,and near infrared region from 225 to 2 500 nm with an interval of 1nm.The change characteristics of soil hyperspectral reflectance curves were studied.The objective of the present paper was to develop a methodology to estimate soil attributes using spectral reflectance.The multiple linear stepwise regression analysis method was used to build hyperspectral models for the prediction of soil attributes,with 22 bands and 13 “reflectance inflexion differences” as independent variables and the soil attributes as dependent variables respectively.Root mean squared error (RMSE) was introduced to test the predictability and precision of the models,and the correlation coefficient was used to evaluate the stability of the models.Some attributes,such pH,SOC,TN,TP,CEC and available N,had the correlation coefficients higher than 0.80,while the value of TK,available P and available K was about 0.68.The results show that the method is feasible to predict the concentration of some soil attributes,while further study should be done for others attributes.