ABSTRACT This study addresses the critical challenge of small sample sizes in deep learning‐based detection of watermelon soluble solids content (SSC) using near‐infrared spectroscopy (NIRS). We propose an auxiliary regression generative adversarial network (ARGAN) that overcomes the limitations of conventional deep convolutional generative adversarial networks (DCGAN) in regression tasks, particularly its poor controllability over chemical properties. The ARGAN framework introduces three key innovations: (1) a semisupervised architecture incorporating continuous SSC labels as conditional inputs to the generator, (2) a discriminative linear activation layer for direct SSC regression, and (3) a mean squared error‐based loss function to enforce chemical‐property correlations. The spectral similarity and SSC distribution consistency generated by ARGAN was higher compared to DCGAN through box‐and‐line plot analysis, average spectral visualization, structural similarity index (SSIMmax = 0.993), maximum mean difference (MMDmin = 0.002), and mahalanobis distance (MDmin = 1.332) analysis. More importantly, ARGAN‐enhanced data improve the prediction performance of partial least squares regression (PLSR), support vector machine (SVM), and convolutional neural network (CNN) models by 9%, 9%, and 13%, respectively, with a significant reduction of 48% in the root mean square error (RMSE) of CNN. As the first successful application of ARGAN for NIRS regression enhancement, this work establishes a novel paradigm for chemically attribute‐controllable data augmentation in agricultural quality detection, effectively addressing small‐sample learning challenges in food spectroscopy analysis.
Li et al. (Sun,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: