Prediction of the Apple Sugar Content of Multiple Varieties by Dielectric Spectroscopy
Yuantao Hua,Yumei Hao,Xiao Zhang
DOI: https://doi.org/10.13031/ja.15624
2024-01-01
Journal of the ASABE
Abstract:Highlights Explore two distinct types of frequency features: optimal frequency and full frequency. Utilize a one-dimensional convolutional neural network (1D-CNN) to classify varieties of apples, employing dielectric spectroscopy features as inputs. Establish models for predicting apple sugar content in various hybrids. Develop an optimal multi-variety search model by integrating a variety recognition model, an individual-variety prediction model, and a classification-first search strategy for accurate soluble solids content (SSC) prediction. Abstract. As a non-destructive detection technology, dielectric spectroscopy encounters challenges when attempting to detect apple sugar content, particularly when aiming to predict sugar content accurately. Variety-specific models may be necessary due to the distinct composition and cell structures present in different apple varieties. Consequently, a single algorithm may prove insufficient for predicting sugar content with precision. Therefore, it becomes imperative to achieve precise sugar content detection across various apple varieties. In the present study, we conducted an analysis of spectral data obtained from apple samples through dielectric spectroscopy. Leveraging these dielectric spectra as input variables, we devised a multi-variety model for predicting apple sugar content. Within this model, the successive projections algorithm (SPA) method was employed to select and optimize the frequency variables. Furthermore, a one-dimensional convolutional neural network (1D-CNN) was applied to the established variety classification model, and the PLS method was employed to develop the individual-variety and hybrid apple sugar content prediction models. When compared to the other two models, the multi-variety model demonstrated superior accuracy in predicting sugar content across various apple varieties. The R 2 p and RMSEP overall prediction sets were 0.915 and 0.00147° Brix, respectively. Meanwhile, the lower values of RMSEP in the other three varieties were 0.0015, 0.0012, and 0.0017° Brix, respectively. This finding indicated that the proposed multi-variety search model was sensitive to variety differences in the prediction of apple sugar content. Therefore, it could improve the accuracy and stability of apple sugar prediction. Keywords: Apple sugar content, Deep learning, Dielectric spectroscopy, Multi-variety search model.