It is very crucial that a representative training set can be extracted from a pool of real samples. Moreover, it is difficult to determine the adapted number of latent variables in PLS regression. For comparison, PLS models were constructed by SPXY, as well as by using the random sampling, duplex and Kennard-Stone methods for selecting a representative subset during the measurement of tangerine leaf. In order to choose correctly the dimension of calibration model, two methods were applied, one of which is leave-one-out cross validation and the other is Monte Carlo cross validation. The results present that the correlation coefficient of the predicted model is 0.996 9, RMSECV is 0.768 1, and RMSEP is 0.736 9, which reveal that SPXY is superior to the other three strategies, and Monte Carlo cross validation can successfully avoid an unnecessary large model, and as a result decreases the risk of over-fitting for the calibration model.