| The country takes agriculture as the main farm and seeds as the first.As agricultural inputs,the quality of seeds has been paid more and more attention,and the traditional seed quality detection methods are difficult to meet the requirements of modern agriculture for nondestructive and rapid seed detection.With the development of computer and spectral technology,hyperspectral imaging technology has been widely used in agricultural detection due to its characteristics of map integration,high efficiency and non-destructiveness.As the oil crop with the largest sown area in China,the breeding of rapeseed varieties is developing towards the direction of "double low and double high",and the quality of rapeseed is also put forward higher requirements in the breeding process.However,at present,the quality detection of rapeseed is usually a chemical method.The operation process is complex and the workload is large,which causes irreversible damage to the seeds.Therefore,how to quickly and efficiently determine the quality of rapeseed is one of the hotspots of current research.In this study,11 different varieties of rapeseed were used as the research object,and the quantitative relationship between oil content,protein content,fatty acid content and spectrum was studied by hyperspectral imaging technology.Nine different pretreatments were carried out on the original spectrum to find the best pretreatment method.On this basis,three different feature extraction methods were used to extract spectral characteristic bands,and three quantitative analysis methods were used to model.The correlation coefficient and root mean square error were used to evaluate the model,and the best inversion model of different quality content was found.The purpose is to provide new method support for hyperspectral technology in seed quality detection.The main results of this study are as follows :(1)In view of the equipment,human and measurement environment and other reasons in the process of imaging spectral data acquisition,there are problems such as redundancy,noise and ambiguity.Mean-Centered,Standardization,Standard Normal Variable,Smoothing,Multiple Scattering Correction,Moving average smoothing,Normalized,2nd-Derivative,1stDerivative 9 spectral pretreatment methods were used to analyze the processed data and the oil content,protein content and fatty acid content of rapeseed with PLS model.The results showed that the multiple scattering correction MSC method performed well in all quality parameters.And determine the subsequent data modeling based on this method.(2)Aiming at the characteristics of multi-band,high dimension and large amount of data in imaging data,in order to improve the modeling accuracy and operation speed.Principal component analysis(PCA),competitive adaptive reweighted sampling(CARS)and continuous projection method(SPA)were used to extract the features,and the quality parameters were used as variables.The number of principal components of oil content,protein content,oleic acid content,linoleic acid content and palmitic acid content extracted by principal component analysis was 5,7,8,6 and 5;the characteristic variables were extracted by competitive adaptive reweighted sampling method,and the characteristic bands were concentrated in 400～415nm and 970～990nm.The characteristic variables screened by continuous projection method are mainly concentrated in 400～415nm and 970～1000nm.(3)Aiming at the problem that the construction method of rapeseed prediction model determines the estimation effect of rapeseed quality,three machine learning methods of support vector machine SVM,least squares support vector machine LSSVM and BP neural network are used to predict the oil content,protein content and fatty acid content of rapeseed.The best prediction model of rapeseed oil content was MSC+PCA+LSSVM,the correlation coefficient Rp of prediction set was 0.912,and the root mean square error RMSEP was 1.324.The best model for protein content prediction was MSC+SPA+LSSVM model,with Rp =0.841 and RMSEP =1.334;the best prediction model of oleic acid content was MSC+CARS+BPNN,the correlation coefficient of prediction set was Rp =0.868,and RMSEP =1.070;in the prediction of linoleic acid content,the results of MSC+SPA+LSSVM model were the best,with correlation coefficient of 0.854 and root mean square error of 1.312.In the prediction analysis of palmitic acid content,MSC+PCA+BPNN model showed the best performance,with correlation coefficient Rp = 0.837 and RMSEP =1.216.The main highlight of this paper is to propose a rapid detection method for quality parameters of rapeseed.This method uses different machine learning quantitative analysis models to invert and predict the oil content,protein content and fatty acid content of rapeseed,so as to avoid damage to rapeseed and determine the quality of rapeseed quickly and accurately,which provides a method support for the use of hyperspectral technology in seed quality detection. |
PDF Full Text Request