Screening naturally mildewed maize kernels based on Raman hyperspectral imaging coupled with machine learning classifiers

Abstract Mildewed maize kernels were common in nature and they were harmful to humans and livestocks. Therefore, it is essential to screen the mildewed maize kernels and improve food security. In this study, Raman hyperspectral imaging technique was used to screen the mildewed maize kernels. Hyperspectral imaging of both embryo‐up and embryo‐down side were taken into consideration and the effect of their weight ratio on the screening accuracy was compared. Pixel‐wise approach and object‐wise approach were both analyzed and different variables selection methods were applied to optimize the input variables. Six machine learning classifiers including random forest (RF), support vector machine (SVM), linear discriminant analysis (LDA), extreme learning machine (ELM), partial least squares discriminant analysis (PLSDA), and naive Bayes classifier (NBC) were used to establish the screening models. The results exhibited 19 variables selected by competitive adaptive reweighted sampling (CARS) method from the weight ratio of 3:7 based on object‐wise approach were the most suitable for establishment of screening model with accuracy of 90.63% in the testing set. The microscopic mechanism of mildewed maize kernels was illustrated by micro‐fluorescence imaging. The study showed that Raman hyperspectral imaging combined with both embryo‐up and embryo‐down sides of maize kernels could be used as effective methods to improve the accuracy of screening mildewed maize kernels. Practical Applications This study provides a comprehensive analysis about the pixels and both sides of maize kernels and gives a nondestructive screening method based on both sides of maize kernels for the mildewed detection. According the mildew of maize kernels, enterprises can effectively improve the storage efficiency of maize kernels to reduce economic losses.