Rapid maize seed vigor classification using deep learning and hyperspectral imaging techniques

• Non-destructive evaluation of physiological characteristics in seeds will transform both the seed testing and seed production industries. • Models based on Deep Learning (DL) were developed for the purpose of classifying maize seed vigor using Hyperspectral Imaging (HSI). • The trained and fine-tuned three-dimensional convolutional neural networks (3DCNN) have demonstrated 100 % sensitivity, 99.89 % specificity, and a Matthews correlation coefficient of 0.9966. • The trained models demonstrate the capability to classify the vigor of maize seeds in a rapid and nondestructive manner. Utilizing conventional methods to assess the seed quality is typically destructive and time-consuming. This research aimed to develop a robust and efficient framework for classifying maize seed vigor using hyperspectral imaging and deep learning. This study involved acquiring hyperspectral imaging data, preprocessing images, and designing convolutional neural network architectures. We explored various network structures, including one-dimensional (1DCNN), two-dimensional (2DCNN), and three-dimensional convolutional neural networks (3DCNN). The impact of wavelength selection on model performance using the successive projection analysis was also explored. The combination of wavelength selection and careful CNN architecture choice significantly contributed to the proposed model's exceptional performance. The 3DCNN model utilizing the full spectral dataset achieved outstanding results, with 100 % sensitivity, 99.89 % specificity, and a Matthews correlation coefficient of 0.9966. Our findings demonstrated the potential of this rapid and non-destructive seed vigor assessment as a practical solution for real-world seed quality control in the agricultural industry.