Near-infrared reflectance spectroscopy was used to address the genetic potential for 990 rice lines for 26 quality traits. The predicted genotypic values for quality traits were calculated using the Mahalanobis distance method and used to measure the genetic similarities among rice varieties. To make the core collection, Manhalanobis distance was employed to calculate the genetic distance among the accessions, and the hierarchical clustering method was used to group the accessions, accompanied by sampling method under the pre-concerted sample proportion (i.e., the ratio of accessions in the core collection to those in the initial population). In this experiment, 24 core collections were developed by using eight hierarchical clustering methods, combined with random, preferred and deviation sampling at a sample proportion of 15% (i.e., 149 for the 15% core collection). These core collections were compared with others constructed at sample proportions of 10% and 20% (i.e., 99 for the 10% core and 198 for the 20% core, respectively). In addition, a trend at increasing sample proportion from 5% to 60% showed that core collection development could be achieved at a sample proportion range of 10–25%. Further results revealed that deviation sampling strategy in combination with the single linkage method retained the greatest degree of genetic diversity of the initial collection. The core collection developed using a sample size of 15% retained the highest degree of diversity, and was stable for all the cluster methods and, hence, the best in developing a core collection of rice quality traits.