Partial least squares regression (PLSR) was employed to build predicting model of the content of soil carbon with on-the-go near-infrared reflectance spectroscopy (NIRS) measurements. The model based on band ratio or normalized difference of NIRS data can improve the prediction precision than the model with the original NIRS data. The reasons might be that the process of band arithmetic combination could reduce the risk of overfitting and it made the model include more useful components and information. The results show that the effective calibration model between field NIRS and the content of soil carbon can be set up by PLSR, and predicting precision can be improved while band arithmetic combination of ratio or normalized difference is performed on the NIRS data before modeling. Thus, it is feasible to estimate the content of soil carbon quickly in the field by on-the-go NIRS measurement.