Detection the quality of pumpkin seeds based on terahertz coupled with convolutional neural network

Pumpkin seeds are nutritious and have some medicinal value. However, the mold and sprouting are produced during the storage of pumpkin seeds. Food safety and quality problems may be caused if they are not removed in time for processing. The traditional testing methods are cumbersome to operate, complex, and destructive in sample preparation. Therefore, terahertz time-domain spectroscopy (THz-TDS) technology was proposed to achieve the detection of the internal quality of pumpkin seeds. Firstly, samples of pumpkin seeds of different qualities were crafted, and they were moldy for 3 days, moldy for 6 days, sprouted and moldy, sprouted and normal pumpkin seeds, respectively. Then, the pumpkin seeds of different qualities were dried, ground, and pressed, and their spectral data were collected. The terahertz spectra of the five types of samples were significantly different. The support vector machine (SVM), random forest (RF), and convolutional neural network (CNN) qualitative discriminant models were established with the raw absorbance spectral data, the preprocessed absorbance spectral data, and the preprocessed and band-screened absorbance spectral data, respectively, where the CNN model based on the raw spectral data has the highest classification accuracy of 96%. The CNN models do not require advance spectral data processing, simplifying the spectral analysis process. And it achieves best classification results in the accuracy of detection compared to traditional chemometric models. The CNN combined with THz-TDS method has great potential for application in the detection of agricultural products. It provides a new detection method for the field of quality detection of agricultural products. In this paper, a terahertz time-domain spectroscopy system is utilized to detect pumpkin seeds. Firstly, the collected terahertz spectral data are preprocessed and band filtered. Then, traditional machine learning methods SVM, RF and deep learning method CNN are used to build the model respectively. Among them, the CNN model built based on the original data has the best classification effect, with a classification accuracy of 96%. The CNN model achieves better classification results compared with the traditional chemometrics model.