Abstract: Adenosine triphosphate (ATP) and its degradation products, which are related to freshness K value, can absorb terahertz (THz) waves due to molecular rotation and vibration and overall vibration of molecular clusters. Thus, THz spectroscopy can be used to detect the K value of pork non-destructively. However, water can strongly absorb THz waves as well, which will affect the accuracy of the obtained results. In this study, different spectral preprocessing methods were compared for their efficiencies in weakening water interference and improving the performance of predictive models in the detection of pork K value by THz spectroscopy. Four spectral preprocessing methods, including multiple scatter correction (MSC), standard normal variate transformation (SNVT), first derivative (FD) and second derivative (SD), were employed to preprocess the original attenuated total reflectance (ATR) infrared spectra. Predictive models were established by back-propagation artificial neural network (BP-ANN) regression algorithm, and their precisions?were compared. The results showed that the FD pretreatment method was the most effective in eliminating baseline drift and improving the spectral quality. Compared with the predictive model without pretreatment, the correlation coefficient of prediction set (Rp) of the FD model was increased from 0.34 to 0.75, and the root mean square error of prediction set (RMSEP) was reduced from 20.24% to 14.36%. This study highlighted the importance of selecting the appropriate spectral pretreatment method to improve the predictive accuracy of models. The BP-ANN model based on FD pretreatment of THz spectra can be used to non-destructively detect pork freshness K value.

Key words: terahertz (THz) spectroscopy, pretreatment, K value, back propagation artificial neural network, non-destructive detection