Data Fusion-Driven Hyperspectral Imaging for Non-Destructive Detection of Single Maize Seed Vigor

Vigor is a vital evaluation index of maize seeds. Accurate and non-destructive detection of single maize seed vigor is of great importance. Hyperspectral images were acquired for the endosperm side and embryo side of individual maize seeds. Four data fusion proposals for both sides of the seed spectra were defined and implemented based on chemometrics methods. Feature engineering methods were used to assist modeling. The data fusion strategies resulted in superior performance for single maize seed vigor detection compared to using only single-side spectra. The convolutional neural network has excellent feature extraction and noise immunity. Feature engineering can further enhance the model performance. Suitable preprocessing algorithms can be used to reduce noise in the original spectra and alleviate the problem of noise amplification by the data fusion approaches. Characteristic wavelength extraction can eliminate redundant information in the original spectra and reduce the parameters of the models. The experiment results demonstrated that the fusion of spectra from both sides of the seed can be successfully used for single maize seed vigor detection and provided a potential method for accurate quality detection of seeds.