Machine learning and experimental testing for prediction of breakage rate of maize kernels based on components contents

Abstract

A high breakage rate (BR) of maize kernels is the main problem during the direct harvest of maize kernels which causes massive grain losses. Thus, a solution was provided for reducing BR in harvest by predicting BR to select a suitable harvest time. The BR prediction models of maize kernels based on moisture, protein and starch contents were studied by using multivariate polynomial regression, stepwise polynomial regression, support vector regression (SVR) and extreme learning regression. SVR with radial basis function (rbf-SVR) was selected for further analysis. The performances of 7 different rbf-SVR models with single and multiple combinations of three components contents were evaluated. The rbf-SVR model constructed with moisture, protein and starch contents (rbf-SVRMs + Pr + St), which were regarded as predictor variables, generated the most accurate BR estimate. The correlation coefficients of the correction set and prediction set were 0.8921 and 0.8776, respectively. The root mean square errors of the correction set and prediction set were 1.3898% and 1.3767%, respectively. The adjusted R2 was 0.7851. The average classification accuracy was 82.17%. As a result, the rbf-SVRMs + Pr + St model can comprehensively evaluate BR and guide the selection of appropriate harvest time, to reduce the BR.