In the identification of haploid maize based on pollen xenia effect, it is of vital importance to determine oil content threshold between haploid and diploid rapidly, accurately and efficiently for automatic classification of haploid by automatic magnetic sorting system in large quantities. A new method based on the least square error was proposed to discriminate haploid from diploid maize seed. Aiming at 10 crosses of maize, through incrementing the number of samples contained in each training set, the size of the training set which can get ideal recognition rate was determined. Principle of least square error used to classify haploid from diploid maize seed was described, and then several experiments were conducted to verify the effectiveness of the proposed method. Seed oil content was measured by nuclear magnetic resonance (NMR) analyzer in the experiments. The least square error method can rapidly determine oil content threshold between haploid and diploid with low amount of samples to realize the practical goal that the haploid correct recognition rate and the diploid correct rejection rate reached more than 90%, and it would provide parameters guidance for the haploid automatic nuclear magnetic resonance (NMR) verification system in the later large-scale classification and improve the efficiency of separation. It was very practical to obtain good classification result with low cost, and it would provide support for the development of haploid engineering.