Establishing Of Phenotype Recognition System Of Anther Dehiscence Status In Cotton Based Deep Learning

Cotton is an important economic crop and strategic material,and an important part of the national economy.However,with the intensification of the greenhouse effect,the frequent occurrence of high temperature weather has seriously restricted the development of the cotton industry.Indehiscence of anthers under high temperature stress are important reasons for the reduction of cotton yield.Therefore,rapid and accurate identification of cotton anther dehiscence is of great significance for identifying important genetic loci and key genes that regulate anther dehiscence under high temperature.The development of computer vision technology and the emergence of big data have promoted the application of deep learning technology in agricultural phenotyping research.Therefore,in this study,a variety of deep learning models(Faster R-CNN,YOLOv5)were proposed to detect the number of anthers and anther dehiscence status,in order to promote the innovation of high temperature tolerance genetics and breeding in cotton.This study achieved the following results:1.After years of multi-point field anther phenotype examination,we obtained38,895 high-definition cotton anther images and selected 2,845 of them as the dataset for training a deep learning-based anther detection model.Firstly,we built a single-stage model detection system based on YOLOv5,and the correlation between the YOLOv5model detection results and the manual count detection results was about 0.6.Meanwhile,the YOLOv5-based anther detection model has a high recognition speed(0.7s per image on a 1050 ti graphics card)and a small model size(only 14 mb),which has the potential for mobile deployment.Therefore,breeding researchers can apply the model to end devices for rapid detection of cotton anther dehiscence condition in the field.2.We trained a two-stage Faster R-CNN model,and the correlation between the detection results of the basic Faster R-CNN model and the detection results of manual counting is about 0.8.Next,three improved strategies(FPN structure,image data enhancement,Multi-Scale module)were proposed for the basic Faster R-CNN model.The improved Faster R-CNN model achieved R~2of 0.8765 in the’open’category,and reached0.8765 in the’close’category.The R~2achieved 0.8539 for the category and 0.8481 for the’all’category.This model has higher detection accuracy than the YOLOv5 model.Using this model,the dehiscence rate of cotton anthers can be quickly and accurately extracted,which can completely replace manual statistics.3.To test the reliability of the improved anther Faster R-CNN recognition model,the anther dehiscence rates of 30 randomly selected cotton varieties under normal and high temperature conditions were counted using the Faster R-CNNN model.The results showed that the anther dehiscence rate of cotton under normal temperature environment was about 84%;the anther dehiscence rate of cotton under high temperature condition was about 35%,and the high temperature adversity significantly reduced the anther dehiscence rate of cotton.Meanwhile,the statistics of manual counting showed that the cotton anther dehiscence rate was about 83%under room temperature environment;about35%under high temperature conditions,which was highly consistent with the detection results of the improved Faster R-CNN model,and the reliability of the cotton anther dehiscence detection model was verified in practice.4.All anther pictures obtained from 510 natural populations of land cotton were examined using the constructed deep learning model,and 35 heat-tolerant cotton germplasm were selected.This is the first time that deep learning technology has been applied to cotton anther dehiscence state recognition,replacing the manual method to quickly and accurately screen out high temperature tolerant cotton varieties,and this model also helps to discover key genes for high temperature tolerance in cotton anthers and promote cotton breeding improvement.