Predicting lodging resistance of maize varieties using leaf hyperspectral imaging

This study aims to investigate the effect of the average spectra of leaf vein, normal reflectance, and whole leaf on the lodging resistance classification of maize varieties. A field trial was conducted in 2018 and 2019, where the hyperspectral image data was collected at the 9-leaf stage for the top leaves of eight maize varieties. The threshold segmentation was used to identify the leaf area, and then the K-means clustering was used to distinguish the leaf into three areas: the normal reflection, dark reflection, and vein area, and finally the average spectral curves were extracted to determine the spectral data characteristics of the lodging resistant samples and the control. The result showed that the average spectral curves of the normal reflectance region and the whole leaf were the typical plant spectral curves, with the distinct absorption bands for blue and red light, whereas, the reflection peaks for green light. However, the whole leaf spectrum contained the spectrum of the dark reflectance region, indicating the lower overall spectral reflectance. There was a low chlorophyll content in the leaf veins. There was almost no absorption band in the spectral curve, where the reflectance in all bands was higher than that in the normal reflectance region, indicating a broader band distribution of the spectral curve. A Kennard Stone sampling was selected to sort the spectral data of each leaf region in each variety. Two parts were divided in the ratio of 3: 1, including the training set and test set. The division of each variety was combined into the final training and test set data to ensure the uniform distribution on each variety. The Max-Relevance and Min-Redundancy (MRMR) feature selection was used to extract the categorical features for each type of leaf area spectra for both lodging resistance and varieties, thereby tranking them in the order of importance. Support Vector Machine (SVM) classification models were built separately for each leaf region using the selected features, where the penalty and kernel parameters of the SVM models were optimized using the grid search to obtain better model predictions. The number of selected features was optimized using the cross-validation method on the training set data. As such, the number of features was selected for the final model considering the prediction accuracy and computational complexity of the model. The experimental results showed that there were significant differences between the spectra of the leaf vein region and the non-vein region of maize leaves. The MRMR was greatly contributed to quickly finding the spectral features that the most associated with the lodging resistance, further improving the prediction of the model. Specifically, the spectral model presented the highest accuracy for the normal reflectance region, with 91.00% and 94.34% predictions for the test sets of the 2018 and 2019 trials, respectively. There were about 35-50 spectral features in each leaf region with the lodging resistance of maize. Among them, there were more features spectra of non-vein areas, compared with the vein areas, indicating more outstanding classification. Therefore, the accurate extraction and the model were greatly contributed to removing the influence of uneven leaf surface on spectral reflection. As such, the higher stability of model prediction was achieved, compared with the average spectrum of the whole leaf. Furthermore, the spectral features of the non-vein region were more suitable for the prediction of the lodging resistance of varieties, compared with the vein region. The spectral model of the normal reflectance region presented a better comprehensive prediction ability for the lodging resistance of varieties. The finding can also provide a strong reference to predict the lodging resistance of varieties using the spectrum of maize leaves. © 2022, Editorial Department of the Transactions of the Chinese Society of Agricultural Engineering. All right reserved.