Research On Intelligent Identification Of Tea Plant Diseases And Pests In Multiple Scenes Based On Spectral And Image Information

Plant diseases and pests pose a major threat to China’s agricultural production and often lead to serious economic losses.The primary factor in the prevention and control of diseases and pests is to conduct many phenotypic detections on plants.On the one hand,it can guide the precise prevention and control of current diseases and pests,and on the other hand,it can screen out genetically mutated plants for subsequent breeding of disease-resistant varieties.At present,the detection of plant diseases,pests and other stresses in China mainly depends on human identification.Although the accuracy is high,it is inefficient and difficult to deal with the large-scale outbreak of diseases and pests.With the development of agricultural informatization in China,technologies such as spectral remote sensing,image processing and machine learning have gradually become effective tools for monitoring plant diseases and pests.However,there are still many challenges in this process,including how to mine effective feature information in massive data,how to reduce the scale of features to improve the efficiency of model application,and how to design suitable plant stress identification models at different scales.At present,tea plant diseases and pests seriously affect the production of tea,but there is a lack of corresponding efficient detection methods.In this paper,three typical and miscible tea plant stresses,namely tea green leafhopper(GL),anthracnose(AH)and leaf sunburn(BR),are taken as examples.Based on spectral and image information,the research on tea plant stresses monitoring methods at leaf,canopy and UAV scales was carried out.The specific research contents include the following:(1)Construction of multi-scale tea plant stresses spectral and image datasets for GL,AH and BR.It mainly includes hyperspectral imaging datasets at the leaf scale and RGB image datasets at the canopy scale.In addition,a larger scale RGB image of tea plant stress is obtained by UAV to evaluate the migration ability of the model and expand the application scenario of the algorithm.(2)In distinguishing three similar stress spectra of GL,AH and BR,an important question is how to extract effective features related to diseases and pests from plant spectra.An ideal spectral feature set should have high sensitivity to target parameters and low information redundancy between features.However,the existing spectral feature selection methods can not meet the above two requirements at the same time.Therefore,this study proposed an algorithm for deep mining plant spectral information,namely the continuous wavelet projections algorithm(CWPA),and tested the algorithm.The results showed that the classification accuracy of CWPA reached 98.08% with only two characteristics in tea plant stresses distinction.In addition,CWPA can also obtain higher model accuracy with fewer features in regression scenarios.Due to the combination of the advantages of continuous wavelet analysis(CWA)and continuous projection algorithm(SPA),the CWPA can find the most sensitive feature set while maintaining the complementarity between features,and has great potential in deep mining of spectral features.(3)Based on hyperspectral imaging technology,a step-by-step detection and discrimination method of tea plant stresses at the leaf scale is proposed.Firstly,the abnormal areas(i.e.the lesion areas corresponding to the three stresses)in the leaves are detected,and then the stress categories are discriminated according to the abnormal areas.In the detection of anomaly area,the exclusive feature set was constructed based on the classical spectral features(differential and continuum removal features),and then the anomaly area detection model is constructed by combining k-means and support vector machine(SVM).The results showed that the overall accuracy of the abnormal area detection model is 92.14%,and the k-means algorithm is used to cluster the leaf images before SVM,which can effectively avoid the problem of irregular edge of scabs in the classification results.Based on the results of abnormal area detection,the CWPA was used to distinguish stresses,and the classification accuracy of the model for three tea plant stresses was 92.01%.The results showed that the hyperspectral imaging information combined with the step-by-step detection and discrimination strategy of tea plant stresses can effectively improve the performance of the algorithm,and has great potential in the detection of plant diseases and pests.(4)At the canopy scale,image data were used to study the methods of detecting,distinguishing and quantifying the degree of multiple stresses in tea garden.A canopy-scale scab identification and segmentation strategy combining deep learning and image processing technology is proposed.Firstly,an optimized Faster RCNN algorithm is used for stress detection.After obtaining the bounding boxes of stress,the scab of tea plant stress was segmented based on the RGReLU feature.Finally,the detection model at the canopy scale is transferred to the UAV image and tested.It is verified that the proposed method can effectively realize the adaptive stress detection at the canopy scale,and can output the scab type and corresponding area ratio.The mAP of the stress detection part reaches 76.07%,and the overall accuracy of the scab segmentation part reaches 88.85%.In addition,the method proposed in this study has strong generalization ability.It can migrate and deploy the model to the UAV shooting scene which is suitable for a wider range of tea garden stress patrol,reduce the data gap between different scenes,improve the utilization of model and data.