Identification Of Rapeseed Variety And Modeling Of Fatty Acid Content Using Hyperspectral Features Fusion

Modern diets pay more and more attention to health,and people have higher and higher requirements for the quality of vegetable oils.The selection and breeding of rape varieties is moving in the direction of"double low and double high"（high oleic acid,high oil content,low erucic acid,low glucosinolate）and yellow seed rape.The detection method of rapeseed in the process of rape variety breeding pursuit of losslessness,convenience,speed,accuracy,etc.is also increasing.In recent years,hyperspectral imaging technology has been widely used in seed variety identification and seed quality parameter detection due to its high resolution,integrated map,high efficiency and losslessness,and large amount of data.This research takes rapeseed as the research object,extracts spectral characteristic parameters and image characteristic parameters based on the map information of hyperspectral imaging,and uses the feature fusion method to establish a hyperspectral recognition model of 11 rapeseed varieties,and the hyperspectral identification of yellow rapeseed The model and the hyperspectral prediction model for the content of four fatty acids,including oleic acid,linoleic acid,palmitic acid and stearic acid,are designed to provide new method support for hyperspectral technology in seed variety identification and quality testing.The main results of this study are:1.In view of the current hyperspectral technology in the field of seed variety identification,there are a large number of spectral band redundancy and sensitive band waste,and the use of multifractal features to reflect the global singularity and effectively deal with the characteristics of non-stationary objects,a hyperspectral feature is proposed.Method for identifying fusion rapeseed varieties.Introduce the idea of feature fusion,construct feature fusion parameters of multifractal feature parameters,image feature parameters and trilateral feature parameters,use index Io to represent the ratio of intraspecies variance and interspecies variance,filter feature parameters through this index,and combine with support vector machines The kernel method（SVMKM）and random forest（RF）method were used to establish a rapeseed variety recognition model.The results show that the best feature fusion parameters are{MF-h（0）,IMF-Si/I1,TRIP-Drmin},and these three parameters form a three-dimensional parameter space,which can identify and identify rapeseed varieties.Recognition of 11 varieties,among which the highest recognition rate between two varieties is 99.67%.When K-fold（Cross-validation）is set to 10,11 rapeseed varieties were identified using random forest（RF）and support vector machine kernel method（SVMKM）,and the average identification accuracy was 92.79%and 92.61%,respectively.2.In view of the complex color changes of the seed coat of Brassica napus L and the inconsistent identification method standards,two identification methods of yellow rapeseed based on hyperspectral characteristics are proposed.The first is the partial least squares（PLSR）regression prediction of R,G and B color channels was carried out by using hyperspectral features.According to the predicted R,G and B values were compared with the RGB color range of yellow and non-yellow seeds determined by international color standards,and the predicted results of yellow and non-yellow seed samples are obtained.The second is the hyperspectral features are used to perform Logit Regression on yellow-seed samples and non-yellow-seed samples,and the recognition rate is calculated according to the Regression results.The results showed that the highest recognition rate of yellow seeds was 96.55%by the PLS method,and the highest recognition rate of non-yellow seeds was 97.78%.Using logistic regression method,the recognition rate of yellow seed reached 98%when spectral index feature fusion parameters were used.3.Aiming at the problems of traditional rapeseed fatty acid determination methods that are time-consuming and labor-intensive,and the need for destructive sampling of seeds,a rapid detection method for rapeseed fatty acid content is proposed.This method is based on hyperspectral imaging technology to establish a characteristic fusion model of rapeseed oleic acid,linoleic acid,palmitic acid,and stearic acid.The results show that:（1）The highest coefficients of determination in the training set of prediction models for oleic acid,linoleic acid,palmitic acid,and stearic acid are 0.8753,0.8893,0.7934,and 0.7824,respectively,and the corresponding parameters are{MF-αmin,IMF-g/（r+b）,SPI-NDSI(R₉₆₃,R₉₉₆)),{MF-αmin,IMF-G/L,SEB-R₄₀₀,SPI-DSI(R₉₇₆,R₉₉₉)},(MF-αmin,SEB-DR₉₉₁,SPI-DSI(R₉₇₆,R₁₀₀₀)}and{MF-h（3）,SEB-R₄₀₀,SPI-RSI(R₄₀₂,R₄₄₄)}.（2）Comparative analysis found that the feature fusion parameters are more accurate and stable than the single parameter model,and have significant advantages overall.In the inversion of four fatty acids including oleic acid,linoleic acid,palmitic acid and stearic acid,the feature fusion is Compared with the single-parameter training set,the coefficient of determination R² of parameters（at least two parameters）is increased by 0.1025,0.113,0.1113,and 0.0953 by 12.65%,12.98%,14.74%,and 12.59%,respectively.However,the results also show that after the feature fusion parameters reach the two feature parameters,the improvement of model accuracy slows down.Therefore,we need to carefully select the feature fusion parameters,which can reduce the workload while improving the accuracy of the model.（3）When the feature parameters of the multifractal method（MF-DFA）were applied to the modeling of rapeseed fatty acids（oleic acid,linoleic acid,palmitic acid and stearic acid）,it was found that the MF feature parameters could greatly improve the accuracy of the combined characteristic model and enhance the generality of the model.Research shows that when the combined feature model contains MF feature parameters,the fitting coefficients of the combined features are improved,Among them,in predicting the bivariate combination parameters of palmitic acid,the training set determination coefficient R² increased by 8.18%,the improvement effect was the most obvious,followed by the bivariate parameter of stearic acid,the improvement rate was 7.2%,and the improvement effect was the worst for linoleic acid With feature combination parameters,the maximum improvement of the training set is only 1.78%.Multifractal features describe the global characteristics of hyperspectral reflectance,which may bring better model performance and generalization for predicting rapeseed oleic acid content.（4）In the oleic acid and linoleic acid inversion model,the accuracy of the model can be improved when the image features are applied to the feature fusion parameters.Among them,the bivariate feature fusion parameters of oleic acid have the best improvement effect,and the training set determination coefficient R² is improved.It is 2.44%,followed by the three-variable characteristic parameters of linoleic acid.The training set determines that the R² coefficient is increased by 1.82%.The main highlights of this paper are the construction of a high-spectral feature fusion non-destructive identification model of rapeseed varieties.The model combines multi-fractal features,image features and spectral trilateral features,and the modeling effect is better than single feature modeling.The second is to propose a method for identifying yellow seeds of rapeseed based on the fusion of hyperspectral features.This method combines the trilateral parameters and the spectral index,which has a good effect on the color identification of seeds.The third is to propose a rapid detection method for rapeseed fatty acid content.This method is characterized by rapeseed hyperspectral multifractal parameters,and inversely predicts the fatty acid composition,which can quickly and accurately predict the fatty acid content.