Deep learning–based multispectral seed classification for detecting adulteration in KDML-105 rice

Abstract

This study aimed to develop and evaluate Deep Learning models for detecting varietal contamination in KDML-105 jasmine rice seeds using multispectral imaging and seed morphological features. Four rice varieties—KDML-105 and three Non–KDML-105 varieties (Chainat-1, RD-15, and RD-6)—were used, with 1,000 seeds collected per variety (4,000 seeds in total). All samples were subjected to imaging, object segmentation, and statistical feature extraction across multiple spectral bands to create inputs for three Deep Learning models: a Bayesian Neural Network (BNN), a Convolutional Neural Network (CNN), and a Recurrent Neural Network (RNN). Experimental results showed that the BNN achieved the highest accuracy on multispectral datasets, reaching 0.99 for both the Physical Reflectance and Multi-Reflectance datasets. The CNN produced slightly lower accuracy, while the RNN consistently performed the poorest across all datasets, indicating that feed-forward architectures are better suited for non-sequential reflectance data. SHAP-based feature analysis identified Min_nir as the most influential variable for classifying both KDML-105 and Non–KDML-105 seeds, followed by Mean_blue and several statistical descriptors from the green and blue spectral bands. These findings highlight the strong discriminative capability of NIR reflectance for separating rice varieties with subtle morphological differences, supported by additional surface-related cues captured in the blue band. Overall, the results demonstrate that Deep Learning combined with multispectral imaging is a viable approach for detecting seed contamination. The identification of key spectral bands suggests the potential for designing a lower-cost inspection system using only NIR and blue channels, enabling the development of automated or real-time sorting solutions suitable for practical implementation in Thailand’s rice industry.