Research On Diagnosis Of Wheat Powdery Mildew Based On The Imaging Hyperspectral Data

Powdery mildew is one of the main diseases of wheat,which seriously restricts the yield and quality of wheat in China.If the disease cannot be found timely and accurately in the early stage,it will easily increase farmers' economic losses,and may cause excessive pesticide spraying to cause farmland environmental pollution.At the same time,the importance of nondestructive monitoring for disease prevention and control is decreasing,but it can provide valuable reference for disaster loss assessment.Therefore,how to diagnose the severity of powdery mildew at the critical stage of wheat growth is very important.Remote sensing technology has many advantages,such as lossless,fast and time-saving,especially the advantage of hyperspectral imaging technology in image and spectral data.It can provide important technical support for target recognition in crop disease and pest monitoring research.In this thesis,wheat powdery mildew was used as the research object,the ImSpector V10E-QE imaging spectrometer was used to collect the target map data,and the combination of computer image processing and machine learning technology was used to predict the disease status of powdery mildew and to classify the disease severity,in order to guide the prevention and control of crop diseases and provide technical support for the assessment of post disaster loss.The main contents and results of this thesis are as follows:(1)The use of computer image processing technology for speckle segmentation,realize the quantitative calculation of Disease Index(DI).First of all,after a median filter to enhance the image,the OTSU segmentation algorithm and fuzzy C-means clustering are used to segment the a and b components in the Lab color space;Secondly,the segmentation of the lesion area was achieved by using the super-red color characteristic 2R-G-B.The experimental results show that:(a)Both of the two algorithms can completely segment the leaf area in the leaf segmentation,but the OTSU segmentation algorithm takes less time and performs more efficiently in actual processing;(b)Take advantage of 2R-G-B can effectively achieve the lesion region segmentation;Then using number of pixels by leaf area and lesion area quantitative calculate disease index.Finally,on the platform of VS2012,the wheat leaf lesion segmentation system was set up to realize the image preprocessing,image segmentation and area calculation of the sample,which provided the data support for the follow-up study.(2)Taking the wheat leaf hyperspectral data of the early stage of disease as the research object,the new vegetation index was calculated and the regression model of early powdery mildew disease was pertinence established.The Relief-F algorithm was used to extract the most sensitive bands and band differences at the early stage of disease.The new vegetation index Powdery Mildew Disease Index(PMDI)was constructed by using the normalized band difference and single-band combination calculation.By analyzing the correlation between disease index and 11 kinds of vegetation index(including PMDI index)and linear model,it was found that PMDI model had the highest coefficient of determination(R2 = 0.8399)and the lowest root mean square error(RMSE = 4.5220),The effect was superior to other vegetation.At the same time,the Normalized Difference Vegetation Index(NDVI)model has the highest coefficient in the common vegetation index.R2 = 0.7771,RMSE = 5.3364.Finally,the Support Vector Regression(SVR)model was established by PMDI and NDVI to construct the early disease index of wheat powdery mildew.The results showed that the PMDI index constructed by the sensitive band screening had better prediction results(R2 = 0.8863 and RMSE = 3.5532),which could provide a method and model support diagnosis of wheat powdery mildew in early stage.(3)The mid-late wheat leaf hyperspectral data were used as the research object to obtain the spectral characteristics of different severity of powdery mildew and to distinguish the different severity grades of late disease.The original normalized spectral data was reduced by using the isomap algorithm.178 samples were used as the experimental data,136 as the training samples and the rest 52 samples as the test samples,and the dimensionality reduction characteristics of the samples were input into the Probabilistic Neural Network(PNN),the overall recognition accuracy of the model can reach 88%,which is higher than the recognition result established by PMDI(overall recognition accuracy is 69%).Meanwhile,for the early sample data of the disease,the prediction based on the dimensionality reduction data the correlation coefficient R2 of model Isomap-SVR is 0.7980,and the root mean square error RMSE is 4.6522.The result of the model is not as good as the PMDI-SVR model.The experiment shows that,for the experimental data in this thesis,all band dates are used to dimensionality reduction is worse than screening a few of the sensitive bands.However,in the middle and late period of wheat disease,the effect of dimension-reduced processing is better than that of the original wave band screening.