Chestnut browning detected with near-infrared spectroscopy and a random-frog algorithm

Key words: 
• cash forestry /
• NIR spectroscopy /
• browning /
• random-frog algorithm /
• least squares-support vector machines /
• partial least squares-linear discriminant analysis /
• chestnut

Abstract: To develop a calibration model for rapid, accurate and nondestructive detection of chestnut browning with peeled chestnut Castanea mollissima by using near infrared spectroscopy technology. Seventy normal chestnuts and 110 browning chestnuts were prepared, and their diffuse reflectance spectrums were collected in the wavelength range from 1 000.00 to 2 500.00 nm. Spectral pretreatment methods, including standard normal variate （SNV）, multiplication scattering correction （MSC）, first derivative （FD）, second derivative （SD）, and detrend, were used and compared first. Then random-frog algorithm was applied to select effective wavelengths （EWs） from the SNV pretreated spectrum. Afterward, a partial least squares-linear discriminant analysis （PLS-LDA） and a least squares-support vector machine （LS-SVM） model were established to classify the browning chestnuts based on EWs, and the results were compared based on sensitivity, specificity and accuracy. For the validation set, the sensitivity, specificity and accuracy obtained by EWs-LS-SVM were 0.92, 1.00 and 95.00%, respectively. The results were better than those of full-PLS-LDA model，full-LS-SVM model and EWs-PLS-LDA model. Also, the random-frog algorithm effectively selected important wavelengths and simplified the discrimination model improving precision and recognition speed. The overall results demonstrate that random-frog algorithm is a powerful tool to select the efficient variables, and EWs-LS-SVM is excellent for the spectral calibration. ［Ch, 4 fig. 3 tab. 22 ref.］