Research On Rapid Detection Method Of Pesticide Residues In Silage Corn Based On Surface-Enhanced Raman Spectroscopy And Hyperspectral

Pesticide residues in corn silage are transmitted and accumulated through the food chain,ultimately endangering human health.Traditional methods of detecting pesticide residues are both time-consuming and expensive,and there has been limited research conducted on rapid detection methods for feed pesticide residues.In this thesis,three pesticides(difenoconazole,carbendazim and acetamiprid),which are commonly used in the growth of corn silage,were investigated for the rapid detection of pesticide residues in corn silage using surface-enhanced Raman spectroscopy(SERS)and hyperspectral techniques,respectively.The goal in this study was to achieve accurate and rapid identification of pesticide components in corn silage.The specific research contents and conclusions were as follows:(1)To investigate the feasibility of surface-enhanced Raman spectroscopy for the identification of pesticide residue species,firstly,a RAPID-785 portable laser Raman spectrometer was used to acquire the Raman spectral information of the samples.Secondly,preprocessing methods such as Savitzky-Golay convolutional smoothing(S-G smoothing)and adaptive iterative reweighted penalized least squares(air PLS)were combined,and finally,a nonlinear discriminant analysis(LDA)-support vector machine(SVM)and convolutional neural network(CNN)pesticide residue identification model was developed.The results showed that,among several modeling methods,S-G-air PLSLDA-SVM modeling was the most effective,with 93.33% recognition accuracy in the training set and 86.67% in the prediction set,but the model accuracy and model running time needed to be improved.To further optimize the model,a competitive adaptive reweighting(CARS)method was used for feature spectrum extraction in the full spectrum interval,and 65 feature variable points were filtered out after 100 sampling cycles.The results showed that the performance of the feature variable-based detection model was better than that of the full-spectrum model,in which the LDA-SVM model with S-G-air PLS preprocessing had the best performance,and the recognition accuracy of the training set of the model was increased from 93.33% to 98.33%,and the recognition accuracy of the prediction set was increased from 86.67% to 93.33%.The results showed that the obtained LDA-SVM model based on 65 feature variables extracted by CARS method could achieve fast and accurate identification of pesticide residues in corn silage with 93.33% accuracy in the prediction set,but the model performance should be further improved.(2)To investigate the feasibility of hyperspectral techniques in pesticide residue species identification,firstly,hyperspectral information of samples was collected by hyperspectral imaging system,secondly,the spectral data were preprocessed by combining S-G and multiple scattering correction(MSC)algorithms,and finally,LDASVM and CNN pesticide residue species identification models were established respectively.The results showed that the MSC-LDA-SVM modeling effect was optimal,and the prediction set identification The accuracy of the prediction set was 83.33%.To further optimize the models,the LDA-SVM and CNN pesticide residue species identification models were developed using CARS and particle swarm(PSO)methods to extract the spectral feature wavelengths.The recognition accuracy of the training set was improved from 96.67% to 98.33%,and the recognition accuracy of the prediction set was improved from 83.33% to 90%.The model running time was shortened from 31 s to 7s,which can achieve fast,non-destructive and more accurate identification of pesticide residues in silage maize raw materials,but the performance was poor compared with the classification model based on surface-enhanced Raman spectroscopy.(3)To further investigate the effect of surface-enhanced Raman spectral data fused with hyperspectral data on the performance of the prediction model,two methods,serial fusion and typical correlation analysis(CCA)fusion,were selected for this thesis.The results showed that the recognition accuracy of LDA-SVM model based on serial fusion of feature spectral segments was 98.33% for the training set and 93.33% for the prediction set.The recognition accuracy of the training set of the LDA-SVM model obtained by CCA fusion based on feature spectral bands was 100%,and the recognition accuracy of the prediction set was 96.67%.It could be seen that the LDA-SVM model obtained by CCA fusion for both spectral feature variables had the best performance,and the prediction set recognition accuracy was significantly improved compared with93.33% for single surface-enhanced Raman spectrum and 90% for single hyperspectrum.Therefore,the LDA-SVM model based on CCA fusion method of surface-enhanced Raman spectra fused with hyperspectrum had a promising future in the detection of pesticide residues in silage maize raw materials.The above results verified the feasibility of surface-enhanced Raman spectroscopy and hyperspectral techniques applied to the identification of pesticide residue species in silage maize raw materials,the effectiveness of recognition model accuracy using feature layer fusion algorithm,which laid the foundation for the development of an efficient and accurate pesticide residue detector for silage maize raw materials and provided a new idea for the rapid detection of pesticide residues in feed.