Research On Non-destructive Detection Of Mango Quality Based On Fluorescence Hyperspectral Imaging Technology

As an agricultural product for rural revitalization in Panzhihua City,Sichuan Province,the local government is trying to use big data,non-destructive detection of agricultural products and other smart agricultural methods to invest in the production of this mango to increase its economic benefits.Among the internal components of mangoes,dry matter content and soluble solids content are important parameters affecting their taste and quality,and these two parameters are closely related to the economic value of mangoes.Therefore,in this thesis,the nondestructive testing of the dry matter content and soluble solids content of mangoes and the quality grading of mangoes will be the main research object.This thesis will combine fluorescence hyperspectral imaging technology and machine learning to nondestructively detect the dry matter content and soluble solids content of mangoes,and also use fluorescence hyperspectral images and deep learning to classify the quality of mangoes,and develop a mango quality grading system based on the mango quality grading model,which will provide theoretical and technical support for the field of rapid nondestructive detection of agricultural products.The main research contents of this thesis include:（1）Non-destructive detection of dry matter content of mangoes.Kite mangoes at the green and ripe stages in Panzhihua were selected as test samples.The MCPLS-SPXY-OSC algorithm was used to pre-process the spectra and achieve orthogonal signal correction.A feature combination extraction strategy was used to improve the stability of the prediction results and reduce covariance.the CARS-RF algorithm is currently the optimal feature extraction combination,which can be derived by comparing the prediction results of three models:support vector machine regression,least squares support vector machine and random forest regression.The best model is the CARS-RF-RFR algorithm,which combines competitive adaptive and random forest regression with a root mean square error of 0.2526and a coefficient of determination of 0.8721 for the training set and a root mean square error of 0.2494 and a coefficient of determination of 0.8775 for the test set.finally,the dry matter of BP＿Adaboost content modeling results were visualized separately,and the comparison showed that the visualization effect of BP＿Adaboost was better,indicating that the method could organically combine the dry matter content and pixel point spatial relationship with the training set coefficient of determination _cR²=0.9944,the test set coefficient of determination R_p²=0.9425.（2）Non-destructive detection of mango soluble solids content.Two methods,GLCM texture features and Spectrum-GLCM texture features,were compared in the model for quantitative detection of soluble solids content of mangoes.Where the Spectrum-GLCM texture fusion is based on a figure fusion strategy,the final 33 optimal feature bands are extracted.Comparing the results of support vector regression（SVR）,least squares support vector machine（LS-SVM）and random forest regression（RFR）for the quantitative prediction of soluble solids,the optimal model was Spectrum-GLCM-random forest regression（Spectrum-GLCM-RFR）,whose training set coefficient of determination and root mean square error were 0.9626|²-²₍（8）|was 0.0008 and|-|was 0.0463.（3）Mango quality grading model and mango quality grading system.Firstly,the original fluorescent hyperspectral images were expanded into a set of 1616 images using eight dataset expansion methods including random flip,horizontal mirroring,vertical mirroring,contrast enhancement,contrast reduction,pepper noise,random panning and distortion scaling to expand the 202-image set into a set of 1616 images.The sample set was then divided into a training set of 1292 images and a test set of 324 images by the SPXY method.The thesis then uses colour space transformation techniques to synthesize pseudo-colour images and combines them with GoogLeNet deep learning modelling.Finally,the t-SNE algorithm was used to visualize the 64 learning channels of GoogLeNet to the 2D plane,discuss the semantic ambiguity in the GoogLeNet model learning process and adjust the GoogLeNet model parameters to achieve an accuracy of 96.89%,a loss value of only 3.11%and a confusion matrix result of 100%for the validation set.The results show that GoogLeNet can be used for mango quality grading,with the advantages of fast recognition and high accuracy,and can be used as the base model of mango quality grading system.The final successful design of mango quality grading system based on this model not only can achieve accurate recognition of mango quality（grade A,B and C）,but also the recognition time is only about 0.2 s.