A Multi-modal Fusion Classification Method for Sorghum Seeds Based on an Improved DenseNet Network

Accurate identification of sorghum seed varieties is critical for ensuring seed quality and promoting sustainable agriculture. However, traditional methods are often timeconsuming and destructive. This study presents a lightweight multi-modal classification framework based on an enhanced DenseNet-121 architecture. By integrating RGB images and hyperspectral reflectance curves into a unified four-channel input, the model effectively captures complementary features from both modalities. It incorporates Inception modules for multi-scale feature extraction and ECA modules for adaptive channel attention to enhance feature discrimination. Experimental results on a sorghum seed dataset show that the proposed model achieves 94.45% accuracy. This approach offers a rapid, non-destructive, and highly accurate solution for seed variety identification, with strong potential in smart agriculture.