A significant problem in genomic selection (GS) is that the number of SNPs is far greater than the number of samples. Eliminating redundant SNP sites is one way to improve the GS model. This paper introduces pairwise squared allele-frequency correlation to evaluate the correlation strength between sites. One of the two sites is randomly removed if the correlation strength exceeds a threshold. Through haplotype analysis, many strongly correlated SNP sites exist in rice, sunflower, and maize. By filtering out 50% of the SNPs through pairwise squared allele-frequency correlation, the prediction accuracy of the GS model does not decrease or even increase slightly. Meanwhile, the training time of the GS model is reduced by half. Through the filtering site method introduced in this paper, the training time of the maize GS model is steadily reduced in six environments. Calculating only the correlation strength between sites on the same chromosome can significantly reduce computational resources, making this method of filtering sites practical. Parallel computing can further reduce the computing time.