In order to realize fast, nondestructive and real-time measurement of parameters (moisture) for pork on-line and divide pork into qualified and unqualified class according to GB/T 18394—2010 (pork moisture limit value ≤76.5%), an on-line detection and classification system was designed and developed based on near infrared reflectance spectra. To determine the detecting distance from the probe to the surface of the sample, optimizing platform was set up, which included spectrum acquisition unit, distance detection unit, light source unit, transmission unit and control unit. NIR spectra (349~1761nm) were obtained from 54 samples with different detecting distances in the range of 5~29mm at approximately 2mm intervals under static conditions. Moisture contents were determined by traditional methods and then related with the spectral information by partial least-squares (PLS) regression models with the first band (349~1435nm), the second band (1037~1761nm) and double-band. The result showed that a detecting distance of 19mm was the best to model. Based on optimal distance, spectra were obtained from 45 samples at static state and on-line by adjusting system of detection distance and multi points simultaneous detection, and then two PLS models were established. The optimal correlation coefficients of the model were 0.915 and 0.906, respectively. The two models results were almost the same. It verified the feasibility of predicting pork moisture on-line. In order to verify stability and precision of models for detecting online, NIR spectra were obtained from another 21 samples. The results showed that the NIR spectral range had an excellent ability to predict the content of moisture (R2=0.8367) in pork online and classification accuracy was 90.48%. Results indicated that NIR spectroscopy was a promising technique to roughly predict moisture of intact fresh pork on-line.