Hyperspectral imaging of beet seed germination prediction

Abstract

The germination prediction of beet seeds is of great significance for sugar beet cultivation. In this study, the near-infrared hyperspectral images of 3072 beet seeds and their corresponding germination states were obtained. We preprocessed the images by cutting, threshold segmentation, dilate, hole-removal, contour extraction, and extracted the average spectrum as the characteristic spectrum. After multi-parameter evaluation and analysis of the six methods, 2D pretreatment was used for mean spectrum. According to the node information gain of the constructed tree model and physical property analysis of beet seeds, 15 characteristic wavelengths were extracted to reduce the dimensionality of data analysis and SVM (RBF), random forest, LightGBM classification models were established respectively. After comprehensive analysis of the prediction effect, LightGBM model was re-established to predict the germination of beet seeds. Analyze external data through testing, founding that the predicted model has a classification prediction accuracy of 89% for the test set. The results show that the germination of beet seeds can be predicted accurately by using the technique of hyperspectral imaging and propose a new idea for batch online non-destructive testing of sugar beet seeds.