Research On Early Detection Approach Of Cotton Verticillium Wilt Combined With Deep Learning And Hyperspectral/Chlorophyll Fluorescence Imaging

Cotton is an important economic crop cultivated worldwide.Xinjiang cotton production is related to the safety of cotton and grain production in China.In the bud stage,cotton is susceptible to Verticillium Wilt caused by Verticillium dahliae.Verticillium Wilt caused a significant loss of cotton biomass,lint yield,and fiber quality.Currently,the detection and control of cotton Verticillium Wilt mainly rely on visual and manual inspections.Visual and manual inspections are time-consuming and laborious,and fungicides prevent and control environmental pollution.This study combined hyperspectral imaging and chlorophyll fluorescence imaging technology to study the phenotypic characteristics of cotton infected with different grades of Verticillium Wilt.This study used neural architecture search to optimize the deep learning architecture of mining cotton phenotypic feature information and constructed a multi-input single-output deep learning architecture based on the fusion strategy.The multi-input and single-output deep learning architecture could deeply mine and amplify the phenotypic changes of cotton infected with Verticillium Wilt at the early stage.Thus,the multi-input and single-output deep learning architecture improved cotton Verticillium Wilt detection and warning timeliness and accuracy.This study provides technical support for effectively reducing the loss of cotton yield,reducing the misuse and abuse of fungicides,achieving accurate control of cost and spraying area,and thus completing the green prevention and control of cotton Verticillium Wilt.The main work and conclusions of this research are as follows:(1)Changes of cotton phenotypic characteristics based on different grades of Verticillium Wilt infection.This study covered conventional varieties of cotton in the market,which were planted in the Verticillium Wilt nursery and naturally infected by Verticillium Wilt.This study combined hyperspectral imaging and chlorophyll fluorescence imaging technology to extract the hyperspectral and chlorophyll fluorescence induction kinetics curves of different grades of cotton leaves infected with Verticillium Wilt.The results of this study showed that the hyperspectral reflectance and chlorophyll fluorescence induction kinetics curves of cotton leaves,which were slightly infected Verticillium Wilt,were slightly different from those of healthy leaves,and the hyperspectral reflectance and chlorophyll fluorescence induction kinetics curves showed a downward and upward trend respectively.The hyperspectral reflectance and chlorophyll fluorescence induction kinetics curves of cotton leaves,which were moderately and severely infected with Verticillium Wilt,showed an obvious upward and downward trend.(2)Early detection of cotton Verticillium Wilt based on single-modal imaging.This study proposed one-dimensional convolution neural networks VGG-1D,Xception-1D,and Resnet-1D,based on Bayesian neural architecture search,in contrast to the classical two-dimensional convolution neural networks(VGG,Xception,and Resnet).The results showed that the constructed one-dimensional convolution neural network(Resnet-1D)had a superior generalization to traditional machine learning(logistic regression,partial least squares discriminant analysis,and support vector machine)in the cotton Verticillium Wilt infection grade detection model.The classification accuracy of Resnet-1D in the training and test sets was higher than 89%,and the cotton Verticillium Wilt early detection accuracy of Resnet-1D was more elevated than 87%.(3)Early detection of cotton Verticillium Wilt based on multi-modal image fusion.The neural architecture search-based multi-input single-output deep learning feature fusion architecture was proposed to detect cotton Verticillium Wilt early,utilizing pixel-level,feature-level,and decision-level feature fusion strategies to combine hyperspectral imaging and chlorophyll fluorescence imaging data.The results of this study showed that compared with single imaging modeling,the traditional machine learning and deep learning using the fusion strategy had improved overall detection accuracy and stability,especially the multi-modal data feature-level fusion deep learning architecture based on Resnet-1D(multi-input single-output deep learning architecture).The detection accuracy of cotton Verticillium Wilt infection level in the training and test sets of the multi-input single-output deep learning architecture was 99.57% and97.88%,respectively.The cotton Verticillium Wilt early detection of the multi-input single-output deep learning architecture was higher than 99%.