Maize Kernel Broken Rate Prediction Using Machine Vision and Machine Learning Algorithms

Abstract

Rapid online detection of broken rate can effectively guide maize harvest with minimal damage to prevent kernel fungal damage. The broken rate prediction model based on machine vision and machine learning algorithms is proposed in this manuscript. A new dataset of high moisture content maize kernel phenotypic features was constructed by extracting seven features (geometric and shape features). Then, the regression model of the kernel (broken and unbroken) weight prediction and the classification model of kernel defect detection were established using the mainstream machine learning algorithm. In this way, the defect rapid identification and accurate weight prediction of broken kernels achieve the purpose of broken rate quantitative detection. The results prove that LGBM (light gradient boosting machine) and RF (random forest) algorithms were suitable for constructing weight prediction models of broken and unbroken kernels, respectively. The r values of the models built by the two algorithms were 0.985 and 0.910, respectively. SVM (support vector machine) algorithms perform well in constructing maize kernel classification models, with more than 95% classification accuracy. A strong linear relationship was observed between the predicted and actual broken rates. Therefore, this method could help to be an accurate, objective, efficient broken rate online detection method for maize harvest.