A total of 156 soil samples with different textures（sand soil(51), clay soil(54) and land soil(51) were collected, and the spectra of all soil samples were scanned with spectrophotometer (ASD FieldSpec3) from 325 to 2500nm. Orthogonal signal correction (OSC) was applied to eliminate the influence of the textures. Soil organic matter (SOM) prediction models of different textural soil samples were then obtained by using partial least square analysis (PLS) and OSC〖CD*2〗PLS. The result showed that when the calibration sample was clay and land soil, the correlation coefficients of PLS and OSC—PLS model were 0.809 and 0.823; when the calibration sample was sand and land soil, the correlation coefficients were 0.837 and 0.734; and when the calibration sample was clay and sand soil, the correlation coefficients were 0.887 and 0.823, respectively. SOM content of another textural soil samples were predicted by using above models, the result showed that the predictive correlation coefficients of PLS and OSC—PLS to sand soil were 0.572 and 0.864; to clay soil were 0.555 and 0.540; and to land soil were 0.643 and 0.721, respectively. The results indicate that OSC can eliminate the influence of texture and improve the prediction precision and solidity of the model.