Detection of drug residues in bean sprouts by hyperspectral imaging combined with 1DCNN with channel attention mechanism

Abstract

As a crucial vegetable food in Asia, bean sprouts are widely cultivated. However, the common practice of using plant growth hormones and other drugs to boost yield and prolong storage poses health risks if not handled properly, so we introduce a novel approach for the non-destructive detection of drug residues in bean sprout cotyledons. In this paper, four experimental groups treated with different drugs and one drug-free control group are tested using hyperspectral imaging (400–1000 nm). A binarization mask is created using the maximum interclass variance approach for the region of interest (ROI) extraction. Preprocessing involves Savitzky-Golay (SG) smoothing and standard normalized variate (SNV). Additionally, three feature extraction methods are used to extract spectral features and reduce the redundancy of the data. The processed data are fed into several traditional machine learning models and latest models, as well as one-dimensional convolutional neural network (1DCNN). The overall accuracy of each model shows the two traditional machine learning models, PSO-SVM and ELM, achieve only 92.3 % and 92.6 % respectively. The 1DCNN model, on the other hand, exhibits an accuracy of 95.5 %, which increased to 96.3 % after applying the channel attention mechanism module. Moreover, in comparison with two state-of-the-art models with accuracies of 95.4 % and 96.5 %, respectively, the proposed model demonstrates competitive performance. The results demonstrate that the 1DCNN model significantly outperforms the two reference machine learning models in this hyperspectral classification task, and the integration of a channel attention mechanism further improves the accuracy, highlighting its potential for application in this field.