Research on multi modal corn seed vitality grading based on three branch cross attention fusion network

Maize is a globally significant crop for both food and feed, and its seed vigor directly impacts germination rate and yield. In the context of intelligent agriculture, there is an urgent need for rapid, non-destructive, and quantifiable methods for assessing seed vigor. This study proposes a novel multimodal fusion approach for maize seed vigor detection, integrating hyperspectral imaging (HSI), electronic nose (ENS), and machine vision (MV) technologies. Multisource data were systematically collected from seeds subjected to varying levels of artificial accelerated aging, thereby constructing a sample set encompassing five distinct vigor levels. Standard germination tests were employed as the ground truth for vigor labeling.Each modality was individually subjected to preprocessing procedures including calibration, denoising, feature extraction, and standardization. The processed data were then fused to construct a comprehensive dataset for model development. Among the unimodal models, classification accuracies of HSI, ENs, and MV reached 91.8%, 94.4%, and 92.0%, respectively. In contrast, the feature fusion network based on three-way cross attention (TCAF-Net) effectively utilizes the complementary information between spectral, olfactory, and morphological features. This model achieved an accuracy of 99.6%, demonstrating superior robustness and stability in distinguishing seed vigor levels.The results validate the efficacy of multimodal data fusion for rapid, non-invasive seed vigor assessment in maize, and provide a promising technical foundation for applications in smart agriculture and seed quality monitoring.