Evidential apple classification model based on combination of Gaussian mass functions

The establishment of stable and highly precise prediction models is of great theoretical and practical significance for the research of the nondestructive testing of the internal quality of fruits with NIR technology. In this paper, focused on Yantai red Fuji apples, probabilistic neural networks (PNN) and partial least squares (PLS) prediction models are developed based on sample data of apple NIR spectra and soluble solid content (SSC). In order to solve the problems of poor adaptability, model instability and uncertainty caused by the hard partitioning of classification in the current research, the two models established in this paper are fused using DS evidence theory. To improve the prediction accuracy of the fusion models, within the framework of DS evidence theory, mass function generation and combination rule selection are discussed. According to the distance between the SSC prediction value and the classification boundary, a Gaussian-based mass function generation approach is designed. Besides, considering Dempster’s rule of combination, Yager’s rule, Sun Quan’s rule and Li-Bicheng’s rule, the effects of combination rules on the apple classification results are investigated. The results show that the classification accuracies of the PLS model and PNN model are 93.9394% and 90.1515% respectively, while the classification accuracy of the fusion model reaches 96.2121%. Compared with our previously proposed linear triangular mass function generation method, the Gaussian-based mass function can better represent the SSC value’s support degree of class prediction. In addition, compared with Dempster’s rule, Dempster’s rule achieves fusion results that are more consistent with the actual situation of apple classification. The proposed combination of Gaussian-based mass function and Li-Bicheng’s rule of combination can not only improve the mass function and make the classification process closer to the practical application, but also increase the prediction accuracy of the fusion model based on DS evidence theory.