Identification of muskmelon seed variety using hyperspectral imaging technology combined with machine learning

Abstract

Abstract: Liquid crystal tunable filter (LCTF) based on hyperspectral imaging technology combined with machine learning is developed to identify muskmelon seed variety rapidly and non-destructively. LCTF-based hyperspectral imaging system equipped with a cold ring LED source is constructed to acquire the reflectance spectra of the muskmelon seeds. Discriminating models based on support vector machine (SVM), linear discriminant analysis (LDA), and convolutional neural network (CNN) are then established to identify the seed variety with reflectance spectra as input. It is found that the LDA model achieved the highest classification accuracy of 100% for the test set while a relatively low value of 96% and 83% was obtained for the SVM and CNN model respectively. To improve classification accuracy of the model, data preprocessing (Savitzky-Golay smoothing and multiple scattering correction) and spectral feature extraction algorithm (successive projections algorithm and principal component analysis) were employed to treat the reflectance data. With these treatments, the classification accuracy of the test set was improved to the highest value of over 99% for the SVM model and 86% for the CNN model. The results showed that the LCTF-based hyperspectral imaging technology combined with machine learning was feasible to identify the muskmelon seed variety. Keywords: hyperspectral imaging; liquid crystal tunable filter; variety identification; muskmelon seed; support vector machine; linear discriminant analysis; convolution neural network DOI:  10.33440/j.ijpaa.20200303.93  Citation: Luo K L, Ye W C, Diao Y H, Zhao J, Chen W J, Liu H C, Long Y B. Identification of muskmelon seed variety using hyperspectral imaging technology combined with machine learning.  Int J Precis Agric Aviat, 2020; 3(3): 14–20.