Total water content is an important quality attribute for consumer satisfaction, and a more accurate pre-detecting method is necessary. The conventional method of partial least squares regression (PLSR) has been widely used in meat water content forecasting. In this study, the cluster analysis of particle swarm optimization algorithm was carried out and calibrated as one of the optimization methods of PLSR with the goal of reducing computation complexity and enhancing the prediction precision. Based on the novel method above, a predicting model of beef water content was developed in wavelength range of 900~2 300 nm, and the best predicting result with Rc=0.920 5 and Rv=0.919 1 was obtained in wavebands of 900~1 400 nm. The samples used in the experiment were beef longissimus collected from the supermarket in that day, and the water contents of samples were detected according to the national standard. Spectra of samples were acquired in reflectance spectral detection system and pre-treated procedure was carried out by means of multiplication scatter correlation method before model construction.