| Rice bacterial blight poses a serious threat to the safe production of rice.Planting disease-resistant varieties can effectively alleviate the widespread spread of disease.However,the loss of disease resistance by the mutation of pathogen are extremely challenging problems in rice disease-resistant breeding.It is necessary to continuously explore new resistance resources.The identification of disease-resistant phenotypes under gene-environment interaction is crucial for determining the expression of resistance genes in breeding for disease resistance.To achieve accurate analysis of resistance phenotypes and efficient mining disease-resistant genes of rice bacterial blight in the field,this study took rice plants with different genotypes as research objects,and used spectral imaging technique,unmanned aerial vehicle(UAV)platform,machine learning and genomics method synthetically to obtain and analyze rice blight resistance phenotype and position resistance genes.This research focuses on expanding the evaluation of disease severity from controlled indoor environment to complex outdoor environment,establishing a universal model for the evalustion of disease severity in the field,tracking the disease severity of different genotypes of rice on a time scale,and realizing high-throughput phenotypic data of time-series UAV combined with rice genome data to mine resistance genes at different growth stages of rice.The specific research contents and results are as follows:(1)A feature selection method based on an attention mechanism was proposed,a disease index was constructed based on the characteristic of image-spectrum merging of spectral images,which realized the expanded application of the research results in controlled indoor environments to complex outdoor environments.In the controlled indoor environment,a model for the evaluation of disease severity based on attentionmechanism(ATT-LSTM)was established using spectral data and image features of visible/near-infrared hyperspectral images of rice leaves with varying disease severity.Important features were selected through the attention mechanism,and a disease index capable of evaluating the disease severity was constructed based on the selected spectral and color features.The results showed that:① The optimal disease severity evaluation results were obtained by the ATT-LSTM model based on the fusion of spectral data and color features with the accuracy of 93%;② The attention mechanism in the ATT-LSTM model assigns attention weights to input variables according to the disease sverity,and can select important features that conform to the law of disease development;③ The constructed disease index has a significant negative correlation with disease severity,with a correlation coefficient of 0.93.The development of disease could be visually displayed by pixel inversion visualization image of the disease index;④ The correlation coefficient between th disease index and disease severity reached 0.84 after expanding it to the data acquired by UAV in the field,surpassing other commonly used spectral indices.(2)The variation rule of rice phenotypic traits under bacterial blight stress was analyzed,and the disease severity evaluation model was established,which provided a method for the evaluation of rice bacterial blight resistance.Based on the phenotypic traits of rice,such as chlorophyll,biomass,water content,and leaf area index,the variation of phenotypic traits of rice with differnet genotypes were studied.The extraction model of phenotypic traits and the evaluation model of disease severity were constructed by UAV multispectral imaging technique combined with machine learning methods.The results showed that:①The self-built neural network could accurately evaluate the variation of rice phenotypic traits under disease stress,with an average Rp2 of 0.80,which showed good estimation and generalization ability;②There was a strong correlation between the SPAD and water content(WC)of rice under disease stress and with the disease sverity with the correlation coefficients were 0.80 and 0.74,respectively.SPAD and WC could represent the development of blight of rice in the field;③The support vector machine model based on spectral data could accurately evaluate the disease severity without increasing the burden of the model(Rp2=0.83,RMSEp=0.70,RPD=2.43).(3)A universal model for evaluating the disease severity of rice bacterial blight was constructed,which solved the problem of incompatibility among different sites and instruments,and provided a solution for realizing the model migration and application.The spectral reflectance under bacterial blight stress was obtained by using the same multispectral camera at different locations and different models at different times.The evaluation model of disease severity across sites/instruments was realized based on transfer learning strategy.The results showed that:① The spectral data from different sites and instruments had distinct feature distributions,making it difficult to apply a disease severity evaluation model based on single feature distribution to other data with different feature distributions.② Migration of 30%of the data from the target domain to the source domain combined with transfer component analysis can effectively improve the application of the disease severity evaluation model between different sites(The coefficient of determination increased from 0.09 to 0.77,and the residual prediction deviation increased from 1.05 to 2.07).③Transfer fine-tuning method realized the migration application of the disease disease severity evaluation model between different instruments.It could obtain reliable prediction results with a small sample collection and avoid feature reduction caused by interdomain data alignment.(4)A disease severity evaluation model on time scale was established,and the feasibility of using UAV phenotypic information for disease resistance gene mining was verified,which provided a research basis for accelerating rice disease resistance breeding.Time-series data at tilling stage,jointing stage and filling stage of rice with different genotype were acquired by UAV based on multispectral camera and RGB camera.The disease severity of rice was tracked over time by fusing the UAV data with the accumulated temperature(AT)data,and the resistance genes were identified together with genome data of rice.The results showed that:① The disease severity could be evaluated based on the UAV data in different growth stages of rice,but the prediction ability varied in different stages.The selection of growth stage influenced the accuracy of the disease severity evaluation.②Accurate evaluation of the disease severity in the field could be achieved by integrating the time-series UAV data with AT data.The SVR model of spectra data fused with AT data fusion obtained the highest prediction ability,with Rp2,RMSEp and RPD of 0.86,0.65 and 2.62,respectively.③Linkage and association studiesshowed that disease resistance at different growth stages was controlled by different genes.The reported resistance gene Xa21 was located on chromosome 11,and a new resistance gene LOC_Os05g44190 was detected on chromosome 5. |
PDF Full Text Request