Application of ensemble neural-network method to integrated sugar content prediction model for citrus fruit using Vis/NIR spectroscopy

Abstract

As consumers' preference for sweetness in food products is increasing, most agricultural product processing complexes currently operate sugar-screening machines. However, every season, the entire prediction models have to be modified to fit the fruit species cultivated that season using a long-drawn process, because separate models are required for each species. Therefore, in this study, species-integrated prediction models were examined based on three citrus species. The species-specific and species-integrated prediction performance of classical partial least squares regression (PLSR)-based, neural-network-based, and ensemble-based models were evaluated. Four different types of ensemble models were proposed depending on the combination method of layers and classification features. The analytical results indicated that the Ensemble Type-4 model exhibited the best performance on both species-specific and species-integrated data, with average 9.9% and 22.1% reduction in RMSETest compared with that of conventional PLSR methods. Furthermore, based on the structural advantages of modularity in ensemble models, effective model maintenance is expected to be possible for future applications in the field.