Abstract:
Abstract: Vigor is a significant indicator for the yield and quality of rice seed in agricultural production and food security. This study aims to fast, non-destructively, and precisely detect the rice seed vigor using hyperspectral imaging and deep learning. Rice seed of Lvhan 1 was selected as the research object. Artificial accelerated aging was also performed on the rice seeds to obtain the samples with different aging degrees. Firstly, a hyperspectral imaging instrument was utilized to collect the hyperspectral images of 1 200 single rice seeds with four aging levels, where the spectral information was extracted from a single sample area. A germination experiment was then carried out, after which all samples were divided into three vigor levels, including high, low, and no vigor. Next, wavelet threshold denoising (WTD), first derivative (FD), and their combination (WTD-FD) were used to preprocess the original spectral data. A model was then established using the preprocessed spectral data. Model evalution was also carried out for the best preprocessing. After that, principal component analysis (PCA) and stack auto encoder (SAE) were adopted to reduce the dimension of spectral data, while extracting spectral features. Thirdly, a detection model of vigor level in rice seed was established using original spectra and the spectral feature data extracted by PCA and SAE. The training and test set were divided ten times to repeat the three Support Vector Machine (SVM) models for higher average accuracies. As such, a better model was determined. GWO was then used to optimize the SVM model, in order to improve the performance of the model. The best model was also determined using the vigor grading model of rice seed. Finally, the best model was used to evaluate the vigor level in a batch of unaged seeds for the generalization performance. The results show that the accuracy of preprocessed spectra (WTD, FD, and WTD-FD) model was higher than that of the original spectra. The WTD-FD accurately reduced the effect of noise and baseline drift in the original spectra, indicating a better pretreatment effect than WTD and FD individually, where the full spectra data was used for subsequent analysis. The better modeling effect was obtained, compared with PCA and original spectra, when the deep features were extracted from full spectra data using SAE. The accuracies of training and test set were 99.08% and 96.47% using the SAE-SVM model, respectively. In grey wolf optimizer (GWO), the accuracies of optimized training and test set in the SAE-GWO-SVM model were 100% and 98.75%, respectively, indicating that improved by 0.92 and 2.28 percentage points, respectively. The accuracy of the model reached 98% for non-aged seeds, approximately to the predictive accuracy of the model (98.75%). Therefore, the hyperspectral imaging combined with the best SAE-GWO-SVM model is feasible for the detection of rice seed vigor in actual production.