Non-destructive classification of rice varieties and mixtures using acoustic absorption spectroscopy and deep learning

Abstract

The widespread issue of adulterating premium rice with cheaper or broken grains significantly diminishes both its nutritional quality and economic value. Traditional methods, such as visual inspection and near-infrared spectroscopy, often prove inadequate when classifying morphologically similar cultivars or analyzing large, bulk quantities. To overcome these limitations, this study introduces and validates a completely non-destructive acoustic methodology for the classification of five major commercial rice cultivars, Jasmine, Basmati, Hashemi, Lenjan, and Broken Lenjan and their intentional mixtures with a lower-grade rice, Anbarbo. We analyzed 1000 bulk samples across four adulteration levels (100, 85:15, 70:30, and 50:50) by measuring sound absorption coefficients (350–1895 Hz) using a four-microphone impedance tube. The raw spectra, following minimal preprocessing, were classified using a customized DeepSpectra convolutional neural network. The DeepSpectra model achieved a robust 84.0% overall accuracy and an F1-score of 0.81 on independent test data, markedly surpassing the performance of both PLS-DA (61%) and a shallow ANN (69%) on pure samples. Interestingly, mid-ratio mixtures (70:30 and 50:50) yielded the highest classification accuracies of 87–89%, which we attribute to the emergence of distinctive spectral signatures. This rapid (under 30 s per sample), low-cost technique is ideal for bulk analysis, providing substantial practical benefits for industrial quality control and verifying the authenticity of granular commodities like rice.