Objective: To improve the precision and accuracy of near-infrared spectroscopy(NIR) calibration model for determination of baicalin content in Xiaochaihu granules. Method: Four NIR characteristic spectrum selection methods were used including competitive adaptive reweighted sampling method(CARS)

monte carlo uninformative variables elimination(MC-UVE)

genetic algorithm(GA) and subwindow permutation analysis(SPA). Model population analysis(MPA) combined with partial least squares(PLS) was used to build five types of models

including CARS-PLS

MCUVE-PLS

GA-PLS

SPA-PLS and all NIR wavelength-PLS. Four model parameters including root mean square error of calibration(RMSEC)

root mean square error of prediction(RMSEP)

prediction mean relative error(PMRE) and the capability of model interpretation for testing set(Q2) were analyzed to evaluate the precision and accuracy of those NIR calibration models. Result: precision and accuracy of the calibration models: CARS > MC-UVE > GA > all NIR wavelength > SPA. The values of RMSEP and decisive coefficients of the calibration models were 1.700 4 and 0.908 7 respectively

whose characteristic spectrum had been selected by CARS. According to the result of paired-t tests at the level of α of 0.05

there was no significant difference between prediction values and measure values in 50 samples. Conclusion: CARS algorithm to screen wavelength can be used to effectively simplify the models

and improve the precision and accuracy of the models

suitable for fast and nondestructive detecting the baicalin content of Xiaochaihu granules.