Hollow discrimination of edamame with pod based on hyperspectral imaging

Abstract

Hollowness is a common defect in edamame with pod that affects its market value and yield. The similar appearance of hollow and normal edamame with pod makes detection and sorting challenging. This study utilized hyperspectral reflectance imaging and transmission imaging to detect hollow edamame with pods. Various classification models were constructed, including partial least squares discriminant analysis, support vector machine, random forest, artificial neural network, and linear discriminant analysis. Six preprocessing methods (SG, MA, MSC, SNV, DT, and WT) and three feature selection methods (SVC-SPA, CARS, and GA) were used to optimize the models. Based on the optimal model, the hollow regions of edamame were visualized, and the proportion of hollow areas was quantified. Based on the optimal threshold for hollow ratio, edamame classification was performed. Results indicate that the SG-SPA-SVM model, derived from hyperspectral transmission data using eight characteristic bands, achieved the best classification performance, attaining a classification accuracy of 98 % at a hollow ratio threshold of 0.48. It offers a scientific basis for quality assessment in related pod and the development of spectrometer applications in production sorting.