Visible near infrared reflectance spectra in the range of 350 nm to 1700 nm were collected from 98 pork samples to develop online, rapid and nondestructive detection system for water content in fresh pork. Median smoothing filter (M-filter), multiplication scatter correlation (MSC) and first derivative (FD) were used as compound preprocessing method to reduce noise present in the original spectrum. Seventy four samples were randomly selected to develop training model and remaining 24 samples were used to test the model. The optimal punishment parameters for the support vector machine (SVM) were determined by using cross-validation and grid-search in the training set. SVM prediction model was developed with the radial basis function (RBF) and the developed model was compared with the model developed by partial least squares regression (PLSR) method. SVM prediction model based on RBF had the correlation coefficient and root mean standard error of 0.96 and 0.32 respectively in the training set. The model obtained correlation coefficient of 0.87 and root mean square error of 0.67 in the test set. The result thus obtained demonstrates the applicability of SVM model for rapid, nondestructive detection of water content in pork.