Improving Prediction Accuracy Of Protein Content In Corn By Using The Multi-Population Genetic Algorithm And Partial Least Squares

Abstract

Although the combination of genetic algorithm and partial least squares (GA-PLS) was successfully used to predict the protein content in corn, there is the problem of premature convergence for the genetic algorithm. In the paper, we replaced the genetic algorithm by the multi-population genetic algorithm to form the combination of multi-population genetic algorithm and partial least squares (MPGA-PLS) to improve the prediction accuracy of the protein content in corn from near infrared spectrum. Based on the public data set which include the protein content and near infrared spectrum for 80 corn samples, we divided the whole near infrared spectrum into 20 intervals, and the infrared spectrum data in 5 intervals were chosen as prediction features by their root mean square errors in the partial least models which transformed infrared spectrum data in each interval into corresponding protein content. Using the infrared spectrum data in the chosen 5 intervals and protein content for the 80 samples, we randomly chose 60 samples for training and 20 samples for testing to compare the GA-PLS and MPGA-PLS by 30 repetitions. Compared with the GA-PLS, the prediction accuracy of MPGA-PLS is significantly improved by 1.88 (P=0.001), and the mean prediction correlation coefficient (CC) is 0.9730. The result illustrates that the proposed MPGA-PLS improved the prediction accuracy of protein content in corn, and may be effectively used into nondestructive testing protein content from near infrared spectrum in many other applications.