Near infrared (NIR) hyperspectral imaging measurement of sugar content in peach was introduced. NIR spectral images (650～ 1 000 nm, resolution: 2 nm) of peach samples were captured with developed hyperspectral imaging setup. Partial least square (PLS) regression prediction model was developed to estimate the sugar content in peach; step-wise backward method was utilized to determine optimal wavelength subsets. Experimental results show that the calibration model with optimal wavelength subsets has a correlation coefficient of prediction of 0.97 and a standard error of prediction of 0.19, the prediction accuracy is higher than the calibration model applied over the whole wavelength, which proves that variable selection plays an important role in improving the prediction accuracy of PLS regression model.