Research on identification of common bean seed vigor based on hyperspectral and deep learning

Abstract

Accurate, rapid and non-destructive identification of common bean seed vigor is of great significance for the planting and efficient utilization of common bean. In this study, five common bean varieties were used as research objects, and four samples with different aging levels were obtained through artificial accelerated aging. Based on the standard germination experiment, the difference in vigor between aged seeds and healthy seeds was verified. Hyperspectral data with aging time of 0d, 2d, 4d and 6d were collected respectively, and one-dimensional average spectra were extracted as modeling datasets using image processing technology. Aiming at the problem of rapid identification of common bean seed vigor, a Multi-scale Spectral Attention Residual Network (MSARN) was proposed in this study. VGG19, MoblieNet, Support Vector Machine (SVM), Random Forest (RF), K-Nearest Neighbor (KNN), and Partial Least Squares Discriminant Analysis (PLS-DA) were used to compare the performance. The results showed that compared to the traditional machine learning models, the deep learning models had better identification results without preprocessing, and MSARN had the best performance. After using the twice Successive Projections Algorithm (SPA), 40 characteristic wavelengths were extracted. The accuracy, precision, recall, and f1-score of SPA-SPA-MSARN for identifying common bean seeds of different vigor levels reached 98.75%, 98.97%, 98.80%, and 98.81%, respectively. Finally, the study applied SPA-SPA-MSARN to five single-variety common bean datasets, and the model was tested to achieve 100% accuracy in identifying vigor levels for four of the variety datasets. This study shows that hyperspectral technology combined with deep learning has great potential in identifying common bean seed vigor.