In this paper, Mengcheng County was taken as the study area in the Huaibei Plain of Anhui Province. A total of 131 topsoil samples (0-20 cm) were collected, their spectral reflectance was measured by Cary 5000 spectrophotometer and the spectral characteristics of typical soil types were analyzed. Moreover, the prediction model of SOM content was established by PLSR. Firstly, whether the different mathematical transformation forms have different effects on the spectral model in predicting SOM content is examined. Secondly, the soil samples were classified according to spectral similarity, and the prediction precision of SOM spectra of different soil types and different spectral classification were compared. The results showed that: 1) The trends of spectral curves of different SOM contents and different soil types were basically similar. The correlation between SOM content and the spectral reflectance was negative, and the curve characteristic varied significantly when SOM content became lower, which is possibly influenced by other factors. 2) PLSR model, established on inverse-log processing of the reflectance, was the best in prediction, with the determination coefficients and the relative percent deviation all increased, the root mean squared error decreased. 3) The precision of SOM spectral prediction model based on the spectral similarity classification was significantly higher than that of the soil type model.