Abstract
INTRODUCTION: Sorghum is an important food and feed crop. Identifying sorghum seed varieties is crucial for ensuring seed quality, improving planting efficiency, and promoting sustainable agricultural development. METHODS: This study proposes a high-precision classification method based on the fusion of RGB images and hyperspectral data, using an improved deep residual convolutional neural network. A spectrogram fusion dataset containing 12,800 seeds from eight sorghum varieties was constructed. The network was enhanced by integrating depthwise separable convolution (DSC) and the Convolutional Block Attention Module (CBAM) into the ResNet50 framework. RESULTS: The CBAM-ResNet50-DSC model demonstrated outstanding performance, achieving a classification accuracy of 94.84%, specificity of 99.20%, recall of 94.39%, precision of 94.52%, and an F1-score of 0.9438 on the fusion dataset. DISCUSSION: These results confirm that the proposed model can accurately and non-destructively classify sorghum seed varieties. The method offers a dependable and efficient approach for seed screening and has practical value in agricultural applications.