Characteristic Information Optimization And Toxin Visualization Of Hyperspectral Data Of Moldy Maize

As an important food crop in China,the safety of maize is closely related to people's lives.Because of the high water content and the large amount of bacteria,fresh maize is easy to be mildewed under high temperature and high humidity conditions.Aflatoxin B₁ and zearalenone are representative toxins produced during the mildew process of maize.Ingestion can lead to liver cell damage under the metabolism of peroxidase in the body.Therefore,the rapid detection and evaluation of moldy maize is necessary.Hyperspectral imaging technology can acquire the spectral and image information of the sample to be tested.The spectral information can detect its physical structure and chemical composition,and the image information can fully reflect its external characteristics.This technique can be used to quickly detect the degree of mildew in moldy maize.However,due to the large amount of information in hyperspectral data,redundancy between information is generated,which is very disadvantageous for the application of hyperspectral technology.In order to reduce the spectral data of moldy maize and eliminate redundant information,this paper proposes a feature wavelength selection method based on continuous projection algorithm fusion information entropy,and explores the correct rate of model construction based on characteristic wavelength,and based on feature wavelength selection.A visualization model of moldy maize toxin was constructed.The specific research work is as follows:1.Set appropriate temperature and humidity conditions and culture 6 maize samples of different mildew grades in a laboratory incubator?50 samples per grade for a total of 300 samples?.The aflatoxin B₁ and zearalenone contents in different mold grade maize samples were determined according to the national standard method.At the same time,hyperspectral images of all samples were acquired using a hyperspectral image acquisition system and corrected in black and white.2.The two preprocessing methods of multivariate scatter correction and standard normal variable transformation are compared.The spectral information and full wavelength information processed by the two methods are substituted into the BP neural network prediction model.The results show that the model constructed by multivariate scatter corrected spectral data has the highest accuracy.The prediction accuracy of the model for aflatoxin B₁ and zearalenone content R²_preand RMSEP were?0.8851,1.0623?and?0.8337,1.3627?,respectively.3.The effective band is determined by the correlation coefficient method.After the initial spectral dimension is used to reduce the original spectral data,the continuous projection algorithm?SPA?algorithm is used to select 8 characteristic wavelengths,and the information entropy principle is used to select the 8 features.The optimal wavelength is selected at the wavelength.Seven invariant moment texture features and six wavelet texture features of the characteristic wavelength image are extracted.The 13 characteristic parameters are used as input parameters of Fisher discriminant analysis?FDA?to obtain each characteristic wavelength?primary selection of 8 characteristic wavelengths?.The grading accuracy rate of the mildewed maize showed that the FDA had the highest grading accuracy rate based on the final selected characteristic wavelength,reaching 99.1%.4.A BP neural network prediction model based on full-wavelength,8characteristic wavelengths and 4 characteristic wavelengths of mildew corn aflatoxin B1 and zearalenone was constructed.The results showed that the prediction accuracy of mildew corn aflatoxin B₁ and zearalenone content was the highest based on the model at 8 characteristic wavelengths.The prediction accuracy of the model for aflatoxin B₁ and zearalenone content R²_preand RMSEP were?0.9769,0.0458?and?0.9841,0.1605?,respectively.?AFB₁ and ZEN?5.Based on the selection of characteristic wavelengths,the partial least squares regression model of moldy maize toxin was constructed.The model was used to predict the content of moldy maize toxin at single pixel,and finally the model of moldy maize toxin was constructed.The results of the thesis can be concluded in three aspects:1)the fusion of spectral information and image information can effectively reflect the information of moldy maize samples;2)the continuous projection algorithm fusion information entropy to select hyperspectral feature wavelength is effective;Visualization of moldy maize toxins can be achieved by predicting the content of moldy maize toxin on a single pixel.