| Chinese cabbage has the reputation of "king of vegetables".After processing and pickling,it can be made into pickles,pickles and so on.It is an indispensable delicacy on the table of ordinary people.Chinese cabbage is widely planted due to its strong adaptability,easy sowing,short growing period and high yield.According to statistics in 2018,the sown area of Chinese cabbage was 20.438 million hectares,and northeast China is one of the main producing areas of Chinese cabbage.Inevitable in the process of vegetable planting,however,the use of pesticides to control plant diseases and insect pests,some farmers lack of medication use common sense and not specification,causing cabbage "poison" and so on the frequent food safety incidents,although the national government at all levels have issued pesticide related policies and regulations,trying to curb the pesticide residues on food safety incidents,but instead,the prohibition of similar incidents.The reasons are as follows:on the one hand,chemical method is the main method for the detection of pesticide residues in Chinese cabbage leaves at the present stage.Although it has high sensitivity and accuracy,it requires a complex pretreatment process before detection,which is time-consuming,laborious,polluting the environment and prone to false negative or false positive problems.On the other hand,the existing pesticide detection methods are relatively backward,and there is a lack of canopy-scale pesticide monitoring methods in the growth of Chinese cabbage.In view of the above reasons,it is of great significance to realize the rapid and nondestructive identification of forbidden pesticide residues in Chinese cabbage at different scales.Hyperspectral technology integrates optics,imaging and microelectronics into one,and has the advantages of high resolution,continuous band and integrated spectrum.Uav remote sensing has the characteristics of low cost,strong operability and high mobility.Federated learning has strong robustness,versatility and scalability,and can undertake complex privacy computing modeling tasks.In this paper,the organic combination of these technologies to the northeast of the common cabbage as the research object,is to identify the target with countries to ban the use of pesticides,to indoor imaging hyperspectral,uav airborne multispectral data as the foundation,from two dimensions of leaf and canopy characteristic spectrum of Chinese cabbage typical disable pesticide extraction,identification,model establishment and the expansion of training,the main contents are as follows:(1)Identification of banned pesticides at leaf scale of Chinese cabbage based on hyperspectrum.Aiming at the shortcomings of traditional chemical detection methods and the hughes phenomenon in hyperspectral data,discrete wavelet transform,hyperspectral technology and convolution neural network algorithm were combined.Firstly,a hyperspectral dimensionality reduction algorithm(DWT)based on discrete wavelet transform was proposed.Secondly,based on the dimensionality reduction algorithm,a hyperspectral identification model of banned pesticides on Chinese cabbage leaf scale was proposed.By comparison test,dimension reduction algorithm of discrete wavelet transform with the traditional competitive adaptive weighting algorithm(CARS),principal component analysis(PCA)dimensionality reduction algorithm,compared to reduce the number of dimensions at the same time,also can better retain the original spectrum curve shape and the relative space position,get disable of pesticide and the corresponding relationship between the characteristic spectrum;After dimensionality reduction,the hyperspectral data were input into convolutional neural network(CNN),multi-layer perceptron(MLP),K-nearest Neighbor algorithm(KNN)and support vector machine(SVM)to establish models and compared.The test showed that the optimal overall accuracy obtained by DWT-CNN was 91.20%,and the Kappa coefficient was 0.89.The combination of hyperspectrum,discrete wavelet transform and convolutional neural network can realize the fast identification of Chinese cabbage leaf scale.(2)Identification of banned pesticides at canopy scale of Chinese cabbage based on multi-spectrum.In order to solve the problem of identification of banned pesticides at canopy scale of Chinese cabbage,discrete wavelet inversion,transfer learning and deep learning algorithms were combined.Firstly,a multispectral dimension raising algorithm(iDWT)based on the inverse discrete wavelet transform was proposed to restore the characteristic spectral information of banned pesticides and solve the existing model utilization problem.Secondly,a canopy scale banned pesticide identification model(iDWT-TL-CNN)based on transfer learning was proposed based on the raised dimension algorithm.Compared with the traditional B-spline interpolation algorithm,the discrete wavelet inversion algorithm can not only achieve better dimensional enhancement effect,but also has a lower algorithm complexity.D after the rise of the multispectral data input existing models,respectively by the model and compare the migration and new learning training,tests show that the training based on the migration of the identification model,and the traditional performance close to the new learning model,discrete wavelet inverse transformation,migration and depth of learning builds canopy scale disable pesticide recognition model,Better model performance can be obtained in the case of fewer data samples and shorter training time.(3)Identification of banned pesticides in Chinese Cabbage based on federal transfer learning.Due to the limitation of the research content and scope,only part of the banned pesticide identification was studied in this study,in order to solve the problems of pesticide identification model expansion,data privacy protection and sharing utilization in the later stage.In this paper,federated learning,transfer learning and deep learning algorithms are organically integrated.Firstly,a dynamic spectral Angle based user selection algorithm(DCSAUS)is proposed to solve the problem of poor federated learning performance under the condition of global data skew.Secondly,a model for identification of banned pesticides in Chinese cabbage based on federal transfer learning(DCSAUS-FTL-CNN)was proposed.Compared with the traditional multi-arm slot machine algorithm,the dynamic spectral Angle user selection algorithm can reduce the impact of data imbalance on the performance of federated learning model,shorten the model convergence time,and improve the model accuracy under the global data skew condition.The identification model of banned pesticides based on federated migration learning can dynamically expand the model and increase the number of recognized pesticides without data sharing,so as to realize the purpose of multiple banned pesticides identification.In this study,the identification of banned pesticides on Chinese cabbage at leaf and canopy scales provided a theoretical basis for rapid and non-destructive detection and identification of banned pesticides on Chinese cabbage at double scales,provided a new regulatory idea for pesticide management in the growth stage of crops,and also provided a new method and technical approach for food safety control. |
