SE-enhanced 1-D CNN with full-band hyperspectral imaging for rapid and accurate maize seed variety classification

Abstract

Introduction: Accurate identification of maize seed varieties is essential for enhancing crop yield and ensuring genetic purity in breeding programs. Methods: This study establishes a non-destructive classification approach based on hyperspectral imaging for discriminating 30 widely cultivated maize varieties from Northwest China. Hyperspectral images were acquired within the 380–1018 nm range, and the embryo region of each seed was selected as the region of interest for spectral extraction. The collected spectra were preprocessed using Savitzky–Golay (SG) smoothing. Several machine learning models—KNN, ELM, and a two-layer convolutional neural network integrated with squeeze-and-excitation (SE) attention modules (CNN2c-SE)—were constructed and compared. Results: Results demonstrated that the CNN2c-SE model utilizing full-spectrum data achieved a superior classification accuracy of 93.89%, significantly outperforming both conventional machine learning models and feature-waveband-based approaches. Discussion: The proposed method offers an effective and efficient tool for high-throughput, non-destructive maize seed variety identification, with promising applications in seed quality control and precision breeding.