Detection for lead pollution level of lettuce leaves based on deep belief network combined with hyperspectral image technology

• Published in: Journal of food safety Vol. 41; no. 1
• Main Authors: Sun, Jun, Cao, Yan, Zhou, Xin, Wu, Minmin, Sun, Yidan, Hu, Yinghui
• Format: Journal Article
• Language: English
• Published: Hoboken, USA John Wiley & Sons, Inc 01.02.2021; Blackwell Publishers Inc
• Subjects: Belief networks; Chemical tests; Embedding; Food safety; Heavy metals; Hyperspectral imaging; Identification methods; Lead poisoning; Lettuce; Model accuracy; Pollution detection; Pollution levels; Stochasticity; Technology; Vegetables
• ISSN: 0149-6085; 1745-4565
• DOI: 10.1111/jfs.12866

Fast detection for heavy metal in vegetables is one of the most important steps to ensure the food safety. A novel method to identify lead pollution levels of lettuce based on hyperspectral image technology was proposed in this study. Firstly, hyperspectral images of lettuce samples cultivated under four lead stress levels (0 mg/L, 50 mg/L, 100 mg/L and 200 mg/L) were collected using hyperspectral image system. Then, a total of 240 spectra were calculated from region of interest (ROI) in the range of 478–978 nm covering 399 bands. Besides, chemical test showed that the excessive level of lead residues content in lettuce leaves were none, slight, moderate and severe. Moreover, conventional models and deep belief network (DBN) were established to determine the best identification model. The discriminant DBN model reached the highest accuracy with the training set of 100% and test set of 96.67%. Finally, t‐distribution stochastic neighbor embedding (t‐SNE) was successful to visualize the feature values in DBN's last hidden layer. This study indicates that it is viable to detect the lead pollution levels of lettuce leaves based on hyperspectral image technology coupled with DBN discriminant model.