Application of Convolutional Neural Network Coupled with Hyperspectral Imaging to Determine the Viability of Peanut Seeds and Its Comparison with Conventional Machine Learning Methods Based on UV-Vis Spectroscopy

Abstract

Seeds can maintain their quality for a limited time; after that, they will lose their germination ability and vigor. Some physiological and physicochemical changes in the structure of the seeds during storage can decrease the quality of the seeds which is known as aging. Therefore, detection of the strong young seeds from the old ones is a vital issue in the modern agriculture. Conventional methods of detection of the seed viability and germination are destructive, time-consuming and costly. In this research, two peanut cultivars, namely North Carolina 2 (NC-2) and Florispan were selected and three artificial aging levels were induced to them. Hyperspectral images (HSI) of the samples were acquired and the seed viability was evaluated using two pre-trained convolutional neural network (CNN) image processing models, AlexNet and VGGNet. The noise of the reflection spectra of the samples was relatively resolved and modified by combining Preprocessing techniques of moving average (MA) and standard normal variate (SNV). Using principal component analysis (PCA), the dimensions were declined and three principal components (PC) were extracted. These PCs were then used as variables in the classification of support vector machine (SVM) and linear discriminant analysis (LDA). The results showed the high capability of CNN architectures such as AlexNet and VGGNet in detection of the seed viability based on the HIS with no pre-processing and feature extraction. The mentioned architectures reached the accuracy of 0.985 and 0.986, respectively. The combination of feature extraction method of PCA with LDA and SVM classifiers showed that the use of a limited number of PCs instead of all wavelengths can decrease the complexity of modeling, while enhancing the efficiency of the models such that LDA and SVM classifiers achieved the accuracy of 0.983 and 0.986 in classification of peanut sees, respectively.

Plant phenotyping relevance

ハイパースペクトル画像とCNN・機械学習を用いてピーナッツ種子の生存性を非破壊的に判定する手法を開発・比較しており、種子状態の取得・推定が研究の中心である。

Code and data availability

The paper's peanut seed hyperspectral images, spectra, and analysis code are not publicly deposited; the Data Availability Statement says datasets are available only from the corresponding author on reasonable request. No public URL or repository is provided.

No evidence-backed public reproduction asset is currently recorded.