Abstract: To establish a model based on near-infrared (NIR) spectra for quickly detecting the freshness of frozen crayfish, NIR spectra of thawed crayfish (tail, meat, and mince) were collected, and data were pretreated by first derivative, multiple scattering correction, wavelet transform (WT), or standard normal transform. The original and pretreated spectral data were correlated to total volatile basic nitrogen (TVB-N) contents using partial least squares (PLS) or convolutional neural network (CNN), and different quantitative prediction models were established and compared. The best model was selected to investigate its accuracy and applicability. The results showed that pretreatment methods had a significant influence on the accuracy of the model, and the CNN model established after spectral preprocessing had a better ability to predict the TVB-N content of crayfish compared with the PLS model. The CNN model based on the WT pretreated spectra of crayfish meat had the highest prediction accuracy for the validation set with correlation coefficients of 0.97 and 0.96, and root mean square errors of 1.26 and 0.93 mg/100 g for the calibration set and validation set, respectively. Moreover, the accuracy, precision, and sensitivity of the NIR method were within reasonable limits, and it had good figures of merit. According to the requirements of fast operation, accurate results, and low damage in practice, the WT-CNN-crayfish meat model was determined as the optimal model for predicting the TVB-N content in frozen crayfish. These results suggested that the WT-CNN-crayfish meat model have a great potential for predicting the TVB-N content and rapidly evaluating the freshness of frozen crayfish.

Key words: near-infrared spectroscopy; crayfish; total volatile basic nitrogen; rapid detection; convolutional neural network; wavelet transform