Rapid And Non-destructive Detection Of Rice Based On Hyperspectral Imaging Technique

China is one of the highest rice production and consumption country,and rice is one of the main food crops in C hina.Due to the diversification of rice producing area,there are many different kinds of rice,and the nutritional quality of different kinds of rice have some differences.With the improvement of people’s life,consumers pay more attention to rice quality and safety.However,traditional laboratory analysises(methods of artificial sense assessment and chemical detection methods)are complicated,lengthy,expensive and difficult for in situ application.Therefore,it is significant to distinct varieties of rice and detect nutrition composition of rice.Because hyperspectral imaging technology contains both spectral and image information,it can not only improve the detection speed,accuracy and stability and lower labor costs,but also can make great contribution to provide rice with high nutritional value and improve exports of C hina’s rice,even can speed up the pace of modern production of the rice industry.In order to detect rice rapidly and non-destructively,this study first applied the spectral information of visible/near-infrared(400 nm-1000 nm)hyperspectral image coupled with spectral data processing methods(and spectra pretreatment methods and optimal wavelengths selection methods)and classification model building methods for classifying the growing areas of rice(central and southern C hina,south C hina,and northeast China),because spectral imformation can represent most internal information of rice.T hen applied the image information of hyperspectral image coupled with texture processing methods to extract the texture information of rice and establish classification models with other image information of rice for classifying the growing areas of rice,because image information can represent most external information of rice.A t last combine the spectral and image information of rice,which contain both internal and external information of rice,to predict the main nutrition components(water,protein and starch)of rice.Specifically,the main research contents and results are shown as follows:1.Rapid and non-destructive classification of rice variety were achieved based on spectral information of visible/near-infrared hyperspectral image.First,training samples and detection samples were selected randomly.Then mean spectra of region of interests(ROIs)of training samples were extracted.Then,classification models of rice were built with spectral data of training rice samples using three building model methods(Partial Least Square Regression(PLSR),Back Propagation Neutral Network(BPNN),and Least Square Support Vector Machines(LS-SVM)),Finally,the LS-SVM model built with original spectra showed the best results(CCR = 92.820%)and therefore used as the optimal method for building models.And then three traditional spectral pre-processing methods((Detrending(DT),Principal Component Analysis(PC A)and Orthogonal Signal Correction(OSC))were respectively used for spectral de-noising,and finally,the LS-SVM model built with OSC spectra showed the best results(CCR = 94.900%)and therefore used as the optima l method for spectral pre-processing.In order to simply the model,the classical wavelength selection method(successive projections algorithm(SPA))was used.Based on the selected wavelengths,optimized model namely OSC-SPA-LS-SVM was established(CCR = 95.360%).Finally,distribution maps of different variety rice were created by transferring the OSC-SPA-LS-SVM model to each pixel in some representative hyperspectral images.2.Rapid and non-destructive classification of rice variety were achieved based on image information of visible/near-infrared hyperspectral image.First,11 characteristic values of shape and 3 characteristic values of transparency of rice sample were extracted.Then,characteristic values of texture of rice were extracted based on different methods(histogram statistics,the length of the gray-level run-length statistics,the gray scale difference statistics and gray gradient co-occurrence matrix statistics).Then,LS-SVM classification models were built based on image information(shape,transparency and texture)of rice sample.Finally,gray gradient co-occurrence matrix statistics was used to extract texture of rice.The best result of LS-SVM classification model built with image data was CCR = 89.741%,which was much lower than the CCR of LS-SVM classification model building with spectra data,therefore,it is not feasible to classify rice varieties based only on image data of rice samples because of the close relation between spectral data and internal information of rice samples.3.Combination of spectra and texture data of hyperspectral imaging for predicting nutrition components(water,protein,amylose,amylopectin,total starch and ration of amylose and amylopectin)of rice were investigated.First,SPA was used to select optimal wavelengths of every nutrition indicator.Then,LS-SVM predicting models of every indicator based on optimal wavelengths spectra were built(R2= 0.913,0.947,0.940,0.846,0.877,0.872).Then,LS-SVM predicting models of every indicator based on 29 image c haracteristic values were built with lower R2.At last,Based on data fusion,LS-SVM predicting models were built with perfect results with high R2(>0.9)for every nutrition indicator,demonstrating that combining spectra with texture data were effective for predicting nutrition indictor of rice.