Hyperspectral RGB Imaging Combined with Deep Learning for Maize Seed Variety Identification

Variety purity is an essential indicator in seed quality detection. Thus, it is necessary to rapidly and non-destructively detect the seed purity. Unlike traditional methods for processing hyperspectral data, this study focuses on computer vision. It aims to reconstruct RGB images from hyperspectral data and employ deep learning techniques to identify the varieties of corn seeds. Firstly, we utilized the diversity of hyperspectral band data to selectively screen the R, G, and B bands with strong feature correlations. These bands were then employed for pseudocolor reconstruction, resulting in a reconstructed dataset with distinct and more interpretive color characteristics than the original dataset. After that, we improved the classic ResNet50 model by adding the coordinate attention (CA) mechanism to the end of each residual block and replacing the global ReLU activation function with SiLU. This improvement enabled the model to capture more precise and detailed features while enhancing its predictive capability. The results showed that without improving the model, the reconstructed dataset generated by the proposed method achieved a classification accuracy of 86.28%, which was a 2.94% improvement over the original RGB dataset and it outperforms 100 randomly generated RGB combinations. While incorporating model improvements, the accuracy of the model on the reconstructed dataset reached 88.18%, surpassing other relevant models and leading to an overall accuracy improvement of 4.79%. The overall study demonstrated that combining hyperspectral image reconstruction with deep learning could be a meaningful tool for identifying and detecting the maize seed variety.