Application of different of models near infrared spectroscopy in detection of apple juice

In order to investigate the stability and application range of near infrared spectroscopy (NIR) models, four calibration models (local model, transferred local model, global model, and optimized global model) were constructed. They were applied to detect three quality parameters (soluble solids content, pH, and electric conductivity) for five varieties of fresh apple juice. The results showed that the prediction accuracy for soluble solids content (SSC) was high (r=0.93, SEP%=3.7%). While electric conductivity was indirectly correlated to NIR, it could hardly be measured by NIR (r=0.84，SEP%=12.7%). The global model of pH had high correlation coefficient (r=0.94), but the differentiation ability of the local model was low (RPD=1.1). The relative standard error of prediction (SEP%), the ratio between standard deviation of quality parameters and SEP were applied to evaluate the stability and application range. The performance of four kinds of NIR models were compared from the aspects of stability, suitability and accuracy. It was found that the application range of model had high influence on the prediction results. Specifically, local model constructed by individual variety samples had high accuracy but poor stability; usually it was only applicable to its own variety and not suitable for other varieties. The method of adding a few samples from the other variety to the current local model so as to construct a new one could broaden its application. The global model constructed with several varieties of samples had high stability although its accuracy was a bit lower than the local model. The method of selecting the representative samples to optimize the global model could reduce the work and the cost for a practical model, and keep its excellent performance as well. Thus it is a worthy method to be recommended.