Non-destructive viability detection of naturally-aged large-sample rice seeds using hyperspectral imaging and deep learning

Precise prediction of rice seed viability is fundamental to enhancing grain yield and ensuring food quality. Traditional viability detection methods are time-consuming, labor-intensive, and seed-consuming, necessitating the development of non-destructive, intelligent, high-throughput detection technologies. Hyperspectral imaging (HSI), as a non-destructive technique, has been widely applied to seed viability detection. The spectral information from different seed components, as well as the number of varieties/germplasms, are key factors influencing detection performance. Therefore, we propose a Mobile Inverted Bottleneck U-Net (MBUNet) for segmenting the combined embryo and endosperm regions in rice hyperspectral images, enabling analysis of how spectral information from different seed components affects model performance. Furthermore, we propose a novel Dual Branch Spectral Transformer Network (DBST-Net) to effectively capture spectral differences between viable and non-viable seeds across 600 naturally aged rice varieties/germplasms, encompassing 58,129 samples. The DBST-Net integrates a multiscale CNN module and a Transformer module to extract local and global spectral features, respectively, while the Multi-Level Feature Fusion Attention (MFFA) module enables deep interaction between these features. Experimental results show that MBUNet achieves a mean Intersection over Union (mIoU) of 95.62 % on the rice seed hyperspectral dataset, outperforming five other segmentation models. Under imbalanced condition of viable and non-viable samples, the DBST-Net achieves the highest classification accuracy of 97.54 % among eight classification methods. The viability-sensitive region analysis reveals that the spectral information of the embryo exhibits the strongest correlation with viability prediction. The transfer experiment demonstrates that the DBST-Net possesses excellent generalization capability. This study provides a novel solution for non-destructive intelligent viability detection of large-sample cross-variety/germplasm rice seeds, offering technical support for safeguarding grain quality.