Selecting key wavelengths of hyperspectral imagine for nondestructive classification of moldy peanuts using ensemble classifier

Abstract

Peanuts can become moldy and produce aflatoxins if transported and stored under improper conditions. Detecting aflatoxin rapidly, non-destructively, and in real-time is important for practical application. This study aimed to identify moldy peanuts by using a small number of key wavelength bands and Ensemble Classifier (EC) based on hyperspectral images. In order to simulate the natural process of fungal infection, the peanuts of three varieties were used to grow mold, and the detailed hyperspectral images of healthy and moldy peanuts (two different degrees) were captured in the 960–2568 nm range. A combination of genetic algorithm and successive projection algorithm was used to select key wavelengths based on the acquired hyperspectral image of peanuts. Following this, an EC consisting of support vector machines (SVM), partial least squares discriminant analysis (PLS-DA), and cluster independent soft pattern classification Classifier (SIMCA) was used to identify healthy and moldy peanuts based on the selected key wavelengths (982 nm, 1180 nm, 1405 nm, 1540 nm, 1871 nm, 1938 nm, 1999 nm). The pixel-wise overall classification accuracy of EC, SVM, PLS-DA, and SIMCA were 97.66%, 97.53%, 95.31%, and 97.36%, respectively. Finally, the kernel-scale classification maps showed the distribution of moldy peanuts amongst the healthy peanuts; this suggests that NIR-HSI is a reliable analytical method for the prediction of moldy peanuts. The overall results support the feasibility of establishing a fast, cost-effective, online multispectral imaging system using a small number of key wavelengths and EC to identify moldy peanuts.