Corn Seed Freezing Damage Identification of Different Sides Based on Hyperspectral Imaging and SPA-2DCOS Fusion Algorithm

In order to improve the utilization efficiency of corn seeds and meet the demand of single-seed seeding technology in agriculture, this study was conducted to explore the effect of freezing damage detection on the endosperm and embryo sides of single corn seeds, based on hyperspectral imaging combined with a feature fusion algorithm and a machine learning method. First, hyperspectral image data of the endosperm and embryo sides of three freezing damage categories of corn seeds were collected, and the average spectra of the endosperm part and embryo part were obtained by threshold segmentation. Then, the spectral data were preprocessed (none, SNV, and 5-3 smoothing), and the feature wavelengths were extracted using the feature wavelength extraction algorithm (SPA and 2DCOS). The modeling accuracy results based on the hyperspectral data of the endosperm and embryo sides at the full waveband and feature wavelength (including feature wavelength fusion) were compared and analyzed. In the endosperm side's freezing damage identification, the SNV+SVM model obtained the highest accuracies of 92.9% and 90.0% with the training set and testing set, based on the full-waveband data. The SNV+SPA-2DCOS+SVM model, based on the feature wavelengths, obtained the highest accuracies of 92.9% and 91.2% with the training set and testing set, respectively. In terms of the embryo side's freezing damage identification, the results on the embryo side were better than those on the endosperm side. The 5-3 smoothing+LDA model, based on the full-waveband data, achieved the highest accuracy results of 97.7% and 95.9% with the training and testing sets. In the meantime, the none+SPA-2DCOS+LDA model, based on the feature wavelengths, achieved the same highest accuracy results with the training and testing sets. When the fusion algorithm consisting of SPA and 2D-COS was used, the model's performance on the endosperm side was better than that of the full-waveband analysis with only 19 feature wavelengths, while the recognition effect on embryo side could be achieved with only 15 feature wavelengths. These results provide a theoretical basis for constructing a multi-spectral detection system for the rapid and nondestructive identification of frozen corn seeds.