The present paper proposed an outlier detection method for spectral analysis based on multi-population elitists shared genetic algorithm. The method was exploited in the NIR data set analysis to remove the outliers from the data set, and partial least squares (PLS) was combined with the proposed method to build a prediction model. In contrast with Monte Carlo cross validation, leave-one-out cross validation, Mahalanobis-distance and traditional genetic algorithm for outlier detection, the prediction residual error sum of squares (PRESS) for moisture prediction model based on the proposed method decreases in the rate of 72.4%, 39.5%, 39.5% and 14.5%; the PRESS value for fat prediction model decreases in the rate of 86.2%, 75.9%, 84.9% and 19.9%; and the PRESS value for protein prediction model decreases in the rate of 56.5%, 35.7%, 35.7% and 18.2% respectively. Results indicated that the method is applicable for spectral outlier detection for different species, and the model based on the data set without the removed outliers is more accurate and robust.