In order to achieve rapid measurement of soil organic matter content, Lou soil was prepared to study soil organic matter determination method based on spectral analysis. Firstly, the soil diffusion reflectance spectrum with range of 900～1700nm was collected by using the portable spectrograph, and the abnormal samples were identified and removed to improve the accuracy. After that, based on the comparison of two different sample dividing methods, the optimal wavelengths of the modeling variables were selected by using successive projections algorithm (SPA). Then, the effect of prediction results of organic matters was analyzed with three linear modeling methods (MLR, PCR, PLS). The results indicated that PLS worked better. The RBF neural network was also built. The results of testing sets showed that the coefficient of determination and root mean square error between measured value and predicted value was 0.8019 and 0.1794 with PLS model, and 0.8281 and 0.1646 with RBF neural network, respectively. Both of them showed a high accuracy which could be used for the fast prediction of the soil organic matter content.