Selecting optimal samples for calibration sets in the visible and near infrared reflectance spectrum prediction of soil organic matter is very important to improve the prediction accuracy of SOM contents. In this paper, the Huangshui river basin in Qinghai Province was selected to screen the optimal sample set method by partial least squares regression model for SOM prediction from visible-near infrared reflectance spectrum. Sample selection methods only considered single soil attribute information such as, concentration gradient method, kennard – stone (KS) method, and sample selection methods based on a variety of soil information including Rank – KS (RKS) method, soil type combined with concentration gradient method and soil type combined with KS. The results showed that under different sample number levels, the optimal sample selection method were obviously different, and the model accuracies of the calibration sample set with multiple soil information, especially the precision of calibration sample set model from KS method combined with soil type with low number of samples, were higher than those of the calibration sample set with single soil information. Adding soil type in soil sample selection can improve the accuracy of model prediction. Under the condition of fixed validation samples and model prediction accuracy, adding soil type into the calibration sample selection can effectively reduce the calibration sample numbers and the prediction cost.