Abstract: Wavelet packet analysis–moving window partial least square(WPA-MWPLS) data processing method was utilized to rapidly detect the caffeine content in tea by near-infrared spectroscopy. Fine multi-level decomposition function of wavelet packet was used to subtract background and noise. At the same time, MWPLS was used to select wavelength ranges that had good relevance with caffeine content in the tea. And then, an adjusted model was established with partial least square. Compared with the model built with the pretreatment of Savitzky–Golay smooth , the prediction correlation coefficient of the model with WPA-MWPLS increased from 0.9170 to 0.9625 and the root mean squared error was reduced from 0.3071 to 0.2463. This method has such advantages as simple pretreatment, few parameters to optimize, and a small number of variables, thus it greatly simplifies the modeling process and enhances the efficiency in building and analyzing the models.