Abstract: In this study, two batches of pork were used as experimental samples, and local partial least squares (LPLS) and dual-band visible and near-infrared spectroscopy were used in conjunction to build a prediction model for the detection of total volatile basic nitrogen (TVB-N). A similarity measurement method based on distance, information measurement, and projection was proposed. A similarity function to evaluate the similarity of different samples was also proposed by weighted summation of Euclidean distance and spectral information divergence-spectral angle (SID-SAM). Then, a similarity factor (SM) was defined and employed to determine the modeling neighborhood window for the establishment of LPLS model. Using the first batch of samples as the modeling basis set, LPLS models for each sample in the second batch of samples were built by optimizing the weights of Euclidean distance and SID-SAM and SM. Compared with the results obtained before establishing the LPLS models, the correlation coefficient (R) for the second batch of samples increased from 0.845 6 to 0.948 1, and the prediction error decreased from 4.581 0 to 2.650 8 mg/100 g, which improved the prediction accuracy of the model. The results showed that using the proposed similarity function and similarity factor, a local space can be dynamically selected in real time according to the spectral characteristics of the samples. The established LPLS prediction models can effectively improve the prediction ability for the validated samples.

Key words: visible and near-infrared spectroscopy, pork, total volatile basic nitrogen, local partial least square