Seedless watermelon seeds (triploid) production is nearly the same as seeded (diploid and tetraploid seeds), although production cost is higher in triploid seeds than seeded one due to the high cost. Many attempts to separate these two kinds of seed before planting have had only slight success with conventional means. In this investigation, we demonstrate the potential of near infrared hyperspectral imaging (NIR-HSI) method to discriminate triploid (3x) seeds from diploid (2x) and tetraploid (4x) seeds, and its application on the online system for real time seed sorting. In order to establish a model for purity discrimination, NIR-hyperspectral images of these three seed groups were collected and analyzed. A multivariate classification model with partial least square discriminant analysis (PLS-DA) was developed, and accuracy analyzed. Over 95% accuracy was obtained in both calibration and validation sets of the PLS-DA model. We further applied the model to the online system for real time discrimination that yielded 80.5% accuracy separation of triploid seeds from its counterparts. The results show good potential of NIR hyperspectral imaging technique for purity separation of watermelon seedless seeds nondestructively, that can be automated for real time sorting for commercial purposes.