Cross-species identification of maize seed storage year by hyperspectral combination of physicochemical parameters

The aging and deterioration of corn seeds are the key factors that affect their vitality. Therefore, identifying the storage time of seeds is of great significance for the preservation of germplasm resources and the quality identification of seeds. This study used machine learning algorithms to identify corn seeds stored for different years by analyzing the correlation between the physicochemical parameters of 103 different corn varieties and hyperspectral bands. Hyperspectral imaging technology was employed to obtain spectral data from both sides of the seed embryonic and endosperm. The near-infrared component analyzer IM9500 and portable ultraviolet-visible fluorescence spectrometer Multi-plex were utilized to acquire the physicochemical parameters of corn seeds. Variance and linear discriminant analyses were conducted to study the trends of physicochemical parameters within corn seeds as storage time increases and their impact on the discrimination of storage time. Physicochemical parameters significantly impacting the discrimination of corn seed storage time were selected. The hyperspectral data underwent black-and-white correction, and the threshold segmentation method effectively separated the background area to obtain the Region of Interest (ROI), with the spectral average of all pixel points on the image serving as the ROI's spectral data. To eliminate interference signals such as background noise, baseline drift, and stray light in the spectral acquisition process, five preprocessing methods Savitzky-Golay (SG) smoothing, standard normal variate transformation (SNV), multiplicative scattering correction (MSC), first derivative (1-Der), and second derivative (2-Der) were applied to the spectral data. Since hyperspectral images contain many bands and redundant information, feature wavelength selection on the complete spectral data was necessary. Competitive adaptive reweighted sampling (CARS) and uninformative variable elimination (UVE) were used for feature wavelength selection. Support vector machine (SVM), backpropagation neural network (BPNN), and convolutional neural network (CNN) classification models were developed. Preprocessing and feature wavelength selection were performed on the spectral data from the embryo and endosperm surfaces. The processed spectral data were used as input for the models, and the classification results of different models were compared. The results indicated that the spectral data from the embryonic side had a better discrimination effect on the storage time of corn seeds. Modeling with preprocessed and feature-selected spectral data significantly outperformed modeling with raw data. By conducting Pearson correlation analysis between physicochemical parameters that significantly affect seed age discrimination and hyperspectral bands, the study associated these parameters with highly correlated feature bands for modeling. The study compared the modeling accuracy of feature wavelength selection, correlation coefficient selection, and the combination of correlation and feature wavelength selection. It verified the model's ability to detect across different corn varieties. The results showed that the BPNN classification model established based on the combination of correlation coefficient selection and feature wavelength selection had the highest accuracy, with a single kernel prediction accuracy of 92.3% and a Colony prediction accuracy of 94.4%. The proposed method has been validated across multiple corn varieties, demonstrating high generalization ability. This study provides a theoretical basis for the precise management of corn seed storage and holds significant practical implications for revitalizing the seed industry.