A portable fiber spectrometer was employed for constructing a transmission spectra detecting system. The system took a 300W tungstenhalogen light source, the integration time for muskmelon samples was 500ms, cumulative integral times of discrete spectra was four, and the smooth method of spectra was boxcar method with four counts smoothing. The system was used to detect the sugar content for 140 samples. Based on the classical least square method (CLS), stepwise multiple linear regress method (SMLR), principal component regression method (PCR) and partial least square method (PLS), four different models between the spectra and the sugar content were established. The test of models indicated that the PLS method was more appropriate for calibration and prediction, the correlation coefficient r was 0.857, root mean square errors of calibration was 0.655, and root mean square errors of prediction was 0.838.