Improving TVB-N prediction in pork using portable spectroscopy with just-in-time learning model updating method

Abstract

Near infrared spectroscopy (NIR) technology is an effective method for nondestructive prediction of total volatile basic nitrogen (TVB-N) in pork. However, the NIR models lack robustness and often fail when used on a new batch. To handle the problem and obtain better prediction performance, a model updating method based on just-in-time learning (JITL) was proposed in this study. A comprehensive similarity criterion considering both input (spectra) and output (TVB-N content) information was designed. Combining a defined similarity factor, the most relevant samples to new batch samples were selected and a local least square support vector machine model was established in real time based on the selected samples. The results showed that the models updated with JITL approach kept a high predictive performance on new independent batch with prediction error decreasing from 2.95 to 1.60 mg/100 g. The robust models made on selected similar samples combined with JITL model updating strategy can support to make NIR spectroscopy a preferred choice for non-destructive assessment of quality features in pork meat.