| Potatoes,as the fourth largest staple food outside rice,wheat and corn,can be processed into staple foods such as bread,noodles,rice flour,etc.,which will promote the rapid development of potato-related industries.During strategy of potato staple food,production is the foundation,processing technology is the means,and product quality testing is the key.Potato quality directly affects the quality of its processed products.At present,the majority of potato quality testing also depends on the traditional chemical methods,these methods are still time-consuming,laborious,destructive samples,polluting the environment and other shortcomings,cannot meet the rapid development of potato staple food needs.In recent years,as a new,green non-destructive testing technology,with multi-band,high resolution,non-destructive,etc.,hyperspectral imaging technology have made some progress in the potato quality nondestructive testing applications.However,the hyperspectral data is a three-dimensional cube data.Its characteristics of redundant information,computational workload,complex processing,running slower,directly affect the modeling speed and prediction accuracy.In order to further improve its detection speed and accuracy,it is necessary to continuously explore and study the data mathematical method of each link in the process of hyperspectral data processing,provide theoretical guidance and technical support for the development of hyperspectral technology,and promote hyperspectral non-destructive testing of potato quality Technology to further promote.In this paper,the potato was used as the research object,and the contents of potato internal components,internal defect category and variety category were used as evaluation indexes.The methods of spectral analysis,stoichiometry,mathematical statistics and data mining were used to improve the quality of potato(Quantitative,qualitative),characteristic wavelength selection and other data processing methods in the detection process to improve the speed and accuracy of the non-destructive testing of potato quality.The main contents and conclusions are as follows:(1)The effects of spectral pretreatment and modeling methods on the model of potato water content,starch,protein and reducing sugar were compared.The principal component regression(PCR),partial least squares regression(PLSR)and support vector machine regression(SVR)models were established for each sample set of potato,and the results were calculated by using themethod of smoothing 13 points,first derivative,second derivative,SNV and decreasing trend transformation,MSC,normalization and orthogonal signal correction and their combination pretreatment method are compared with the modeling results of the original spectrum.The results showed that the optimal detection model of water content,starch,protein and reducing sugar content of potato were PLSR model.The optimal detection model of water content was the result of pretreatment of the orthogonal signal correction and the extraction of 8 principal components,the R 2,Rp2 of the calibration set and validation set are 0.7948 and 0.7870,respectively.TheRMSEC and RMSEP of the calibration set and the validation set are 0.3882% and 0.3735%respectively.The optimal detection model of starch and reducing sugar content was obtained when the spectrum was smoothed by 13 points,and the starch was obtained when the number of principal components was 12,the R 2,Rp2 RMSEC and RMSEP of the optimal starch contentdetection model were 0.8312?0.8286?0.4498% and 0.3986%,respectively.The R 2,Rp2 RMSECand RMSEP of the optimal reducing sugar starch content detection model were 0.8516?0.8464?0.0729% and 0.0758%,respectively.The optimal detection model of protein content was obtained by the method of multi-scattering correction.The number of extracted principal components was15,and the R 2,Rp2,RMSEC And RMSEP of the model were 0.7919,0.7904,0.0456%,0.0414%respectively.(2)The characteristic wavelength selection method of the quantitative analysis model of each component in potato was studied.Genetic algorithm(GA),non-information variable elimination(UVE),competitive adaptive weighting algorithm(CARS),continuous projection algorithm(SPA)and random frog algorithm(Random-frog)were used to analyze the effects of the ability of the sample hyperspectral characteristic wavelength selection.It is determined that the Random-frog algorithm has the best screening result for the hyperspectral characteristic wavelength of potato samples.(3)The identification model of black heart disease in potato was studied.The effects of different spectral pretreatment methods and pattern recognition methods on the qualitative analysis of black potato and qualified potato were compared.The original spectral spectra were pretreated by Gaussian smoothing,moving smoothing,SG smoothing,MSC,first derivative,second derivative,variable normalization,orthogonal signal correction,de-trend transformation and its combination.Different identification models of Spectral information of potato black heart disease were established.The results show that the BPNN and PLSLDA models established by SG derivative method and orthogonal signal correction combined pretreatment spectrum have better recognition effect on potato black heart disease,and the correct rate of sample recognition is96.84%.(4)The characteristic wavelength selection method of potato black heart disease identification model was studied.BPNN and PLSLDA recognition model were established by random-frog,sub-window rearrangement analysis(SPA)and interval effect analysis(MIA).The experimental results show that the recognition accuracy of the model is 99.29%,92% and 97.37%,respectively.The correctness of the model is 99.29%,92% and 97.37% respectively.The model uses 45 wavelength variables,about 1/5 of the total wavelength variable,and the number of hidden layer nodes is 6.(5)The potato variety identification model was studied.The effects of different spectral pretreatment methods and pattern recognition methods on the identification of potato varieties were compared with the three varieties of Eugene 885,early white and Zhongshu No.5 of Keshan potato.The experimental results show that the DA model and BPNN model are better when the spectrum is preconditioned,and the overall recognition rate of the model is 98.15%.(6)The characteristic wavelength selection method of potato variety identification model was studied.Non-information variable elimination(UVE),competitive adaptive weighting algorithm(CARS),genetic algorithm(GA),Random-frog algorithm and genetic algorithm combined with continuous projection algorithm(GA-SPA)were used to carry out spectral characteristics wavelength optimization,respectively,to establish DA and BPNN identification model.The experimental results show that the GA-SPA-BPNN model can improve the recognition performance of potato varieties to a certain extent.The correctness of the model,the validation set and the overall sample are 100%,95.24% and 98.77% respectively. |
PDF Full Text Request