Rapid And Nondestructive Detection For Maize Aging Based On Hyperspectral Imaging Technology

The deterioration of maize quality is due to a decrease in nutritional quality and an increase in the content of harmful substances during the ageing process.This is why it is essential to test maize for freshness and ageing.Traditional chemical detection methods are time-consuming,contaminating,damaging to the sample itself and cumbersome.In this study,a rapid and non-destructive method for the detection of maize ageing was established by hyperspectral imaging technology,which can solve these problems very well.In the experiment,three different varieties of maize,Deydan 5,Yuyu 30 and Zhengdan 958,were used as the research subjects and subjected to accelerated ageing treatment at high temperature(40±0.5°C)and high humidity(90%±3 RH)to study the trends of peroxidase(CAT)activity,peroxidase(POD)activity,fatty acid values and electrical conductivity of maize.Seven pre-treatment methods(SG,SG1,SG2,D1,D2,SNV,MSC),two characteristic wavelength selection methods(CARS,SPA)and two classifiers(BP,RF)combined with chemometrics were also used to develop quantitative prediction models for each of these four indicators,and the best quantitative prediction model was selected using hyperspectral imaging.By rescanning the spectral images of three different varieties of maize seeds at 0,70,140 and 210 days,a square area of 230 * 230 = 52,900 pixel points was selected as the region of interest for spectral data extraction and prediction,and the distribution of each indicator was visualised based on each indicator under the corresponding optimal prediction model.The establishment of a non-destructive and rapid method for the detection of maize neoteny has led to the following main findings:(1)CAT activity of three different varieties of maize kernels(Deyuan 5,Yuyu 30 and Zhengdan 958)all showed a highly significant negative correlation with ageing time,with the CAT activity of Yuyu 30 decreasing more sharply during ageing,by 93.39%;the CAT activity of Zhengdan 958 decreasing second,by 84.58%;and the CAT activity of Deyuan 5decreasing the least,by POD activity was significantly negatively correlated with ageing time,with the most drastic decrease of 98.46% for Zhengdan 958,the second most drastic decrease of 93.48% for Deidan 5 and the slowest decrease of 83.92% for Yuyu 30.The fatty acid values were all positively correlated with the ageing time,and during the ageing period,the fatty acid value of Deyuan 5 increased the most,reaching 169.08%;the fatty acid values of Yuyu 30 and Zhengdan 958 increased less,at 102.94% and 116.27% respectively.The conductivity also showed a significant positive correlation with ageing time,with Yuyu 30 showing the sharpest increase in conductivity,with an increase of 1174.30%;Deyuan 5showing a more dramatic increase in conductivity,with an increase of 760.29%;and Zhengdan 958 showing the slowest increase in conductivity,with an increase of 528.23%.(2)Based on the performance analysis of the CAT active full-band classification model,the D1-RF prediction model had the best performance,where the coefficient of determination of the test set RP2 = 0.9573 and the root mean square error RMSEP was5.2053.Based on the performance analysis of the CARS characteristic band classification model,the SG1-CARS-RF prediction model had the best performance,where the coefficient of determination of the test set RP2 = 0.9610,with a root mean square error RMSEP of5.3814,in which a total of 53 wavelengths were extracted,accounting for 20.70% of the total wavelengths.Based on the performance analysis of the SPA feature band classification model,the SG1-SPA-RF prediction model performed best with a coefficient of determination RP2= 0.9660 and a root mean square error RMSEP of 5.1338 for the test set,of which a total of 23 wavelengths were extracted,accounting for only 8.98% of the total number of wavelengths.Visualisation of the distribution of CAT activity under the best prediction model SG1-SPA-RF was achieved.(3)Based on the performance analysis of the POD activity full-band classification model,the MSC-RF prediction model performed best,where the coefficient of determination RP2 = 0.8504 and the root mean square error RMSEP was 49.8500 for the test set.based on the performance analysis of the CARS feature band classification model,the SG1-CARS-RF prediction model performed best,where the coefficient of determination RP2 = 0.9090 and a root mean square error RMSEP of 47.7322,of which a total of 31 wavelengths were extracted,accounting for 12.11% of the total wavelengths.Based on the performance analysis of the SPA feature band classification model,the SG2-CARS-RF prediction model performed best with a coefficient of determination RP2= 0.8212 and a root mean square error RMSEP of 54.4754 for the test set,of which a total of 10 wavelengths were extracted,accounting for only 3.91% of the total number of wavelengths.A visual distribution of POD activity under the best prediction model SG1-CARS-RF was achieved.(4)Performance analysis of the full band classification model based on fatty acid values showed the best performance of the SG2-RF prediction model,where the coefficient of determination RP2 = 0.9567 and root mean square error RMSEP of 3.8408 for the test set.performance analysis of the classification model based on the CARS feature band showed the best performance of the SG1-CARS-RF prediction model,where the coefficient of determination RP2 = 0.9504,with a root mean square error RMSEP of 4.1216,and a total of38 wavelengths extracted,accounting for 14.84% of the total wavelengths.Based on the performance analysis of the SPA feature band classification model,the SG2-SPA-RF prediction model performed best,where the coefficient of determination RP2= 0.9655 and the root mean square error RMSEP was 3.6255 for the test set,with 31 wavelengths extracted,accounting for only 12.11% of the total number of wavelengths.Visualisation of the distribution of fatty acid values under the best prediction model SG2-SPA-RF was achieved.(5)Based on the performance analysis of the conductivity full-band classification model,the SG1-RF prediction model performed best,where the coefficient of determination RP2 = 0.9386 and the root mean square error RMSEP was 3.3363 for the test set.based on the performance analysis of the CARS feature band classification model,the D1-CARS-RF prediction model performed best,where the coefficient of determination RP2 = 0.9502,with a root mean square error RMSEP of 3.2784,and a total of 53 wavelengths extracted,accounting for 20.70% of the total wavelengths.Based on the performance analysis of the SPA feature band classification model,the SG1-SPA-RF prediction model performed the best,where the coefficient of determination RP2= 0.9507 and the root mean square error RMSEP was 3.3789 for the test set,with a total of 35 wavelengths extracted,accounting for only 13.67% of the total number of wavelengths.The visualisation of the conductivity distribution under the best prediction model SG1-SPA-RF was achieved.