Peach quality prediction using dielectric spectroscopy and generative adversarial networks

Dielectric spectroscopy is a promising technology in fruit quality detection. However, its effectiveness is often hindered by limited sample sizes. To address this, peaches were used as samples, and their firmness, soluble solids content (SSC), and moisture content were predicted using dielectric spectroscopy integrated with generative adversarial network (GAN). Firstly, dielectric spectra (20–4500 MHz) and quality parameters were collected from 160 ‘Shahong’ and ‘Qianzhouhong’ peaches (intact and pulp). Then, conditional generative adversarial network (CGAN) and conditional deep convolutional generative adversarial network (CDCGAN) were applied to augment the training set. Finally, the augmented training set was used to develop prediction models using partial least squares regression (PLSR), support vector regression (SVR), and one-dimensional convolutional neural network (1D-CNN). Results indicated that CDCGAN provided superior data augmentation compared to CGAN. The 1D-CNN trained on the CDCGAN-augmented training set achieved the best performance, with the root mean square error of the prediction set for intact peaches decreasing by 16.547%–38.846% compared to the original PLSR model, while those for pulp samples decreased by 27.500%–52.686%. These findings show that GAN-based data augmentation effectively enhances dielectric spectroscopy accuracy for predicting peach internal quality parameters, providing a novel approach for nondestructive fruit quality assessment.