To make Vis/NIR diffuse transmittance technique applied to quality prediction for watermelon in motion, the dynamic spectra detecting system was rebuilt. Spectra detecting experiments were conducted and the effects of noises caused by motion on spectra were analyzed. Then the least-square filtering method and Norris differential filtering method were adopted to eliminate the effects of noise on spectra smoothing, and statistical models between the spectra and soluble solids content were developed using partial least square method. The performance of different models was assessed in terms of correlation coefficients (r) of validation set of samples, root mean square errors of calibration (RMSEC) and root mean square errors of prediction (RMSEP). Calibration and prediction results indicated that Norris differential method was an effective method to smooth spectra and improve calibration and prediction results, especially, with r of 0.895, RMSEC of 0.549, and RMSEP of 0.760 for the calibration and prediction result of the first derivative spectra.