To determine the moisture content in de-enzyming green tea leaves rapidly and nondestructively, prediction models were established based on visible-near infrared spectroscopy. Diffuse reflection spectra of 192 samples were collected with a portable field spectrometer (FieldSpec 3, ASD), among which 144 samples were partitioned to a calibration set and 48 samples to a prediction set using the sample set partitioning method based on joint X-Y distance. 11 sensitive bands were selected with correlation coefficient method, and then moisture content models of partial least squares and principal component regression, artificial neural network and their combination were established with the preprocessing methods of the first derivative and moving average filter. The model comparison showed that the prediction model of partial least squares regression was the best when 5 principal components were adopted. The calibration and prediction correlation coefficients were 0.990 and 0.819 respectively, and the root mean square errors of calibration and prediction were 0.011 and 0.037 respectively, and the mean error of predicted moisture content was 3.30%. 