Application case of HSI in quantitative detection of foodborne pathogenic bacteria: quantitative detection of S. aureus contamination on mutton using 2D-COS combined with 1D-CNN-LSTM-Self-Attention

The feasibility of in-situ quantitative detection of foodborne pathogenic bacteria in meat products using short-wave infrared hyperspectral imaging technology (SWIR-HSI) was explored by taking the detection of Staphylococcus aureus ( S. aureus ) content on mutton as an example. Meanwhile, a method for extracting characteristic wavelengths with strong interpretability was put forward to remove redundant information. Based on this, combined with a deep integration modeling approach, the feasibility of establishing a high-accuracy spectral detection model using fewer wavelengths was demonstrated. Specifically, 1D-CNN regression model was established firstly, and the Self-Attention mechanism was introduced to solve the overfitting problem of the model and improve its accuracy. Then, methods of 2D-COS, MRMR, GA and CARS were used for extracting obtain characteristic wavelengths. Next, the LSTM module was added to the 1D-CNN-Self-Attention to construct 1D-CNN-LSTM-Self-Attention model, and the simplified models were established based on characteristic wavelengths. The results showed the model established based on the 7 characteristic wavelengths selected by 2D-COS had best performance, and the R 2 , RMSE and RPD of test set for the model were 0.8911, 0.6601 and 3.04, respectively. It indicated that the combination of 2D-COS with 1D-CNN-LSTM-Self-Attention could effectively detect the concentration of S. aureus on mutton. The results provide theoretical support and reference for developing multi-spectral detection instrument for foodborne pathogenic bacteria on meat.