Identification of wheat seeds from bran layer using optical microscopy and deep learning

Purpose: This study aims to automate the identification of grain varieties and select the most suitable wheat genotypes for specific ecological conditions using Artificial Intelligence (AI)-based systems. The goal is to facilitate high-yield and high-quality production through pre-sowing analysis. Method: Seeds from nine wheat genotypes with different qualities were used, and cross-sections of the wheat genotypes were photographed under a light microscope to create a specialized dataset. A Convolutional Neural Network (CNN)-based automated wheat identification framework was then proposed, utilizing both shallow and deep architectures. Findings: The experiments confirm that CNN-based methods are highly effective in extracting distinctive features from wheat bran and accurately identifying wheat seed varieties. Conclusion: The research successfully distinguished nine varieties and found that a simpler model (ResNet18) outperformed deeper networks, offering a practical solution for agricultural verification. Keywords: wheat;classification;optical microscopy;deep learning;seed analysis