Construction of an Intelligent Detection System for Rice Seed Viability Based on Deep Learning and Hyperspectral Imaging

Traditional rice seed vigor detection methods have limitations of low efficiency or strong destructiveness, which are difficult to meet the needs of large-scale detection in modern seed industry. To address this, this study constructed a non-destructive detection system for rice seed vigor based on improved ResNet50SE and hyperspectral imaging. The system collects spectral and spatial texture data of 5 main rice varieties in Sichuan through a visible-near-infrared hyperspectral imaging platform ($\mathbf{4 0 0 - 1 0 0 0} \text{nm})$. After radiation correction, wavelet denoising, and feature wavelength extraction using the Successive Projection Algorithm (SPA), the fused spectral-spatial features are input into the improved ResNet50-SE model (introducing the SE channel attention mechanism to enhance the weight of key features) for vigor grading. Experimental results show that the system's average accuracy for grading high/middle/low vigor seeds reaches $\mathbf{9 2. 3 \%}$, which is $\mathbf{1 5. 6 \%}-\mathbf{2 2. 3 \%}$ higher than traditional machine learning methods (such as SVM and Random Forest) and 3.2% higher than the basic ResNet50 model. This research provides an intelligent technology for rapid and non-destructive detection of rice seed vigor.