Detection Of Tomato Diseases And Insects Based On THz-NIR Hyperspectral Fusion

China is one of the largest tomato producing and consuming countries in the world.Tomato leaves are the main part of diseases and insect pests.Tomato leaves infected by diseases and insect pests not only gradually change their external physiological morphology,but also slowly destroy the internal structural components of tomato leaves.At present,the detection of tomato pest leaves mainly depends on manual work,which has the disadvantages of time-consuming,laborious and low measurement accuracy.In view of the above problems,this study comprehensively uses the technologies of image processing,near infrared(NIR)hyperspectral,terahertz time domain spectroscopy and multi-source information fusion,and takes the greenhouse crop tomato as the research object to study the rapid detection means and recognition methods of tomato leaf mold and tomato leaf miner under the environment of agricultural facilities,so as to provide reference for computer image information processing It provides a reference for the application of spectral information analysis system and fusion technology in modern agriculture.The specific research contents and results are as follows:(1)The data sets of tomato leaf mold samples and tomato leaf miner samples were collected,and the infringed levels were divided by image processing technology.According to the obtained image information,the disease and insect spots on the image features are obtained by image recognition technology,and the leaf area occupied by the disease spots is calculated,which is used as the basis for classification(level 0,level 1,level 2 and level 3).(2)The detection method of tomato diseases and insect pests based on NIR hyperspectral was proposed.By comparing the effects of savitzky Golay smoothing(SG)combined with first derivative(FD),variable Standardization(SNV)and multivariate scattering correction(MSc)algorithms on the preprocessing of raw near-infrared hyperspectral data,the preprocessing algorithm of SG smoothing combined with SNV has the best effect,greatly eliminates the influence of noise,and is suitable for the preprocessing of raw spectra of tomato leaf mold samples and tomato leaf miner samples.The joint interval partial least squares algorithm(sipls)and genetic algorithm(GA)were used to screen the characteristic wavelengths of tomato leaf mold samples and tomato leaf miner samples.Back propagation neural network(BPNN)and k-nearest neighbor(KNN)modeling methods were used to establish the recognition model between healthy samples,tomato leaf mold samples and tomato leaf miner samples.Ga-knn model had the best recognition effect on tomato leaf mold samples.The recognition rates of grade 0,grade 1,grade 2 and grade 3 samples were 100.00%,95.24%,88.88% and 90.91% respectively,and the overall recognition rate was 92.65%.GA-BPNN model has the best recognition effect on tomato leaf miner samples,and the recognition rate and overall recognition rate of samples at all levels are 93.33%.(3)A detection method of tomato diseases and insect pests based on terahertz time domain spectroscopy was proposed.The spectral data of absorbance dimension and power spectrum dimension are extracted from terahertz time domain spectral images,and the influence of noise is removed by SG smoothing combined with SNV preprocessing algorithm.Principal component analysis(PCA)and uninformative variable elimination(uve)were used to screen the characteristic wavelengths of absorbance and power spectrum dimensions of tomato leaf mold samples and tomato leaf miner samples.BPNN and KNN modeling methods were used to establish the recognition model between healthy samples,tomato leaf mold samples and tomato leaf miner samples in the dimensions of absorbance and power spectrum.Under the absorbance dimension,pca-knn model has the best recognition effect on tomato leaf mold samples.The recognition rates of grade 0,grade 1,grade 2 and grade 3 samples are 100%,95%,100% and 93.33% respectively,and the overall recognition rate is96.67%.Pca-bpnn model and pca-knn model have the best recognition effect on tomato leaf miner samples in the dimension of power spectrum.The prediction accuracy of the two models at all levels is 100%,93.33%,93.33% and 86.67%,and the overall prediction accuracy is 93.33%.(4)An improved pest detection method based on multi-source information fusion is proposed.Based on the proportion of lesion area,near-infrared hyperspectral data and time-domain terahertz spectral data,an improved diagnosis model of tomato leaf health state was established by introducing a priori probability table and auxiliary variables and optimizing Bayesian network.By fusing three kinds of prior information,the posterior distribution of tomato leaf health parameter changes was recalculated,and the estimation results of health parameter changes were greatly improved.Finally,the improved Bayesian network has a recognition rate of 97.12%for tomato leaf mold samples and 93.35% for tomato leaf miner samples.