| Rice is an important food crop in china,and its stable and high-quality output is crucial to ensuring the basic food supply and maintaining national food security.However,rice bacterial blight seriously endangers the growth of rice and reduce its the yield and quality.Cultivating disease-resistant varieties is the most cost-effective and environmental-friendly way to control the disease.Breeding disease-resistant cultivar relies on the rice disease-resistant phenotypes to determining the effectiveness of genes.In traditional research,the acquisition of rice phenotype primarily relies on manual observation,which has lots of shortcomings such as low efficiency,large errors and strong subjectivity.And there is an absence of high-efficient analyszing methods for complicated phenotype features.Spectral imaging technology which is applied in plant phenotyping,is able to non-destructively and rapidly characterize both the internal and external information.Nowadays,deep learning algorithm also reveals its excellent ability in feature analysis.In order to evaluate the phenotype of disease-resistant rice more efficiently,accurately and objectively,assist breeding experts to intelligently discriminate disease-resistant rice varieties,a method system for obtaining and analyzing the phenotyping of bacterial blight resistant rice were established.The specific research contents and conclusions are as follows:(1)Based on spectral imaging technology and deep learning,the discriminant model of rice seed genotype was established to realize accurate identification of rice seeds with bacterial blight resistance.Near-infrared hyperspectral imaging and terahertz spectral imaging technology combined with self-built neural network,were applied to establish disease-resistant rice seed discrimination models based on the average spectrums and the spectral images,namely Seed Net-1D and Seed Net-2D,respectively.A fine-tuned learning method based on Seed Net-1D was proposed for discriminating rice seeds in different harvest years.The feature analysis process of deep learning was visualized using feature dimensionality reduction algorithms.It was shown that:(1)based on terahertz spectrums,the discrimination accuracy of Seed Net-1D and Seed Net-2D reached 91.49%and 83.51%,which surpassed the traditional machine learning models and deep learning models based on near-infrared spectrums;(2)Transfer fine-tuning method improved the discrimination accuracy of disease-resistant rice seeds in different haverst years reducing the cost of collecting new samples;(3)The deep learning model effectively promoted intra-class cluster and inter-class separation.(2)Combined with multiple imaging technology and detection methods,the effects of bacterial blight on the physiological and biochemical changes of rice genotypes were studied,which provided a theoretical basis for evaluation of rice bacterial blight resistance.Under the bacterial blight strees,by using transmission electron microscopy,the variation of rice leaf microstructures was captured,and the content of physiological and biochemical substances in rice leaves were detected based on UV spectrophotometry.And based on fluorescence quantitative polymerase chain reaction,the relative expression levels of defense genes of different rice genotypes were explored.It was shown that:(1)the bacterial blight pathogenic bacteria in the vascular bundles of disease-resistant rice leaf was few and scattered,and the cell structure was complete and plump;(2)The changes of photosynthetic pigment,malondialdehyde and glutathione in leaves at different time after infection revealed differences in stress response in disease-resistant rice;(3)The relative expression levels of disease-related defense genes in resistant rice were all up-regulated,while the expression of disease-related defense genes in disease-susceptible rice was inhibited.(3)In the controlled indoor environment,a diagnostic model of rice leaf disease status and bacterial blight lesion proportion was established based on spectral imaging technology and attention mechanism,achieving an accurate assessment of rice bacterial blight resistance.Chlorophyll fluorescence imaging technology were used to preliminarily explore the separability of bacterial blight resistance,and a discrimination model of infection status and disease resistance was established.Based on the visible near infrared hyperspectral technology and the standard for dividing bacterial blight resistance,a model,namely LPnet,with dual-branch structure of lesion proportion prediction and infection state diagnosis was established.It was shown that:(1)disease-resistant rice was able to maximally alleviate the disruption of photochemical reaction under the influence of bacterial blight pathogens;(2)The disease diagnosis model based on rice chlorophyll fluorescence characteristics had proved the feasibility of applying spectral features to evaluate rice bacterial blight resistance;(3)The lesion proportion regression R~2 and infection state discrimination accuracy of LPnet reached 0.96,and 92.39%;(4)The correlation coefficient between the spectral index formed by extracting the LPnet attention weight and the lesion proportion was as high as 0.97.By combining the spectral index and LPnet,the disease level of rice leaves was predicted with an accuracy of 85.41%.(4)In the complex outdoor environment,lesion proportion caculation model was established by using image instance segmentation algorithm,which realized the accurate evaluation of rice bacterial blight resistance.Relying on RGB imaging technology,images of rice leaves in complex outdoor environment were collected.An imporved Mask R-CNN model which was utilized to segment the rice leaf and lesion among the RGB images was established by combining the attention mechanism of convolutional module and Tversky loss.The effects of rice leaf images,which were collected in complex outdoor environment in different years and scenes,on the performance of the improved Mask R-CNN model were studied,and the disease level of rice bacterial blight was calculated according to the segmentation results.It was shown that:(1)there were differences in the spread speed of leaf lesion between the disease-susceptible and disease-resistant rice;(2)The average intersection ratio of leaves and lesions of improved Mask R-CNN based on the adaptive histogram equalization algorithm reached 93.38%and 78.64%,respectively;(3)The regreesion R~2 between the combined results of the segmentation masks and the actual lesion proportion reached up to 0.97,and the disease level discrimination accuracy rate was also as high as 90.90%.(5)By integrating the discriminant regression models and the phenotype analysis methods,the software of rice bacterial blight resistance phenotype analysis system was designed and built.The interface of the analysis system was designed and produced by using Py Qt5.The whole system included three main branch structures,including the identification of disease-resistant rice seeds,the evaluation of rice bacterial blight resistance based on spectral images,and the RGB-based branch of rice bacterial blight resistance evaluation.According to the module design of the branch structure,multiple specific functions were built and the system was packaged into executable software by Pyinstaller,which provided a basic platform for the application of the methods proposed in this paper. |
PDF Full Text Request