Detection Method Of Drought Stress Response In Barley Seedling Stage Based On Hyperspectral Image

Barley is the fifth cereal crop in the world,with edible,feed and industrial brewing value.It is of great significance for ensuring global food security and social and economic development.Using hyperspectral spectral data containing multidimensional spectral information and the relationship between barley drought response,barley drought resistance can be quickly and non-destructively identified,and feature selection is a prerequisite for using hyperspectral data for trait investigation.In view of this,this study constructed a drought stress dataset using PEG-treated barley seedling leaves,and segmented the high-resolution hyperspectral images of barley leaves to obtain their specific positions in the images.Based on chi-square test,elastic net method and extremely random tree method were used for secondary feature extraction,and two feature datasets were obtained.On this basis,several binary classification models were compared on different feature sets.The main results are as follows:1.According to the Mask-Rcnn model,about 300 images were annotated,and the head layer was trained on the basis of the model pre-trained with the coco dataset.The final mask accuracy was 84.6%,and only a small part of the leaf tip area was predicted inaccurately in the predicted image.This confirms that the method has good transferability.2.Based on chi-square test,elastic net method and extremely random tree method were used to perform secondary feature extraction on the mask data obtained by Mask-Rcnn model segmentation,resulting in two feature datasets,reducing the spectral features from 462(392.34nm-1009.43nm)bands to 40 bands,achieving feature sparsification.The bands selected by elastic net method are roughly in the areas of 420 nm,500nm,550 nm,630nm,700 nm,750nm,and the bands selected by extremely random tree method are roughly in the areas of 500 nm,550nm,620 nm,640nm,700 nm,720nm.3.After validation by multiple models,this paper believes that compared with the dataset selected by the extremely random tree model,the feature set selected by the elastic net feature selection has better test results on models such as Logistic Regression(LR),K-Nearest Neighbor(KNN),Random Forest(RF),CNN.Under the elastic net dataset,KNN model performed the best with an accuracy of 96.02%,and Gaussian model performed poorly with an accuracy of 77.67%;under the extremely random forest tree dataset,KNN model also achieved the best accuracy of 94.84%.Similarly,Gaussian model performed the worst with an accuracy of 78.77%.KNN model showed a large difference in performance on these two datasets,which was 1.18%.In generalization validation,both KNN and CNN models had some degree of overfitting,and CNN model performed better.This study is based on the hyperspectral information of barley after drought stress,and uses elastic net method and extremely random tree method for feature selection,effectively reducing the data dimension and improving the model performance.The study screened out the spectral features related to drought stress,providing the corresponding basis for rapid detection of barley drought stress response phenotypic information extraction,and providing new clues and means for crop drought monitoring.The results of this study are helpful to develop a crop drought monitoring system based on hyperspectral technology,and provide scientific basis for agricultural disaster prevention and control.