Research On Identification Of Soybean Varieties Based On Hyperspectral Image Technology

In China, in many agricultural products, soybean plays an important role, an d it is the most important source of vegetable oil and an important supply source of plant protein. Soybean is one of the favorite ingredients of the people. In recent years, soybean demand continuing increasing, soybean imports continuing rising, scientific and rapid and accurate detection of soybean quality become more important. Including oil, protein, the nutritional value quite different in among soybean varieties, so it can not be ignored that species identification. Hyperspectral reflectance data of 10 soybean seed varieties of 100 grains was collected as study object, to realize the non-destructive identification of soybean seed varieties. The study confirmed that hyperspectral image technology is feasible for non-destructive identification of soybean seed varieties. Image feature chose in this study is good at characterization of soybean. It would be helpful to realize dynamic real time non-destructive identification of soybean varieties in terms of theoretical basis and technical support.The main research contents of this paper are as follows:In this study, hyperspectral reflectance data of 10 soybean seed varieties of 100 grains was collected in spectral region of 400.918-999.53 nm, including hyperspectral reflectance spectra information and hyperspectral reflectance image respectively used as characterization samples. Compare the identification ability with mean or standard difference spectra or both them as inputs when hyperspectral information as characterization samples. When image feature information as characterization samples, four characteristics parameters were extracted from three feature bands images which contain the most variety information. The four characteristics parameters are energy, entropy, angular second moment and correlation based on gray level co-occurrence matrix(GLCM).Hyperspectral reflectance image denoised by median smoothing, correction of multiplicative scatter or normalization. Compare the identification ability with different pretreatment in the same model when hyperspectral information as characterization samples. Hyperspectral information as characterization samples, median filter smoothing was chosed.Four characteristics parameters of three feature bands images were extracted, including the mean and standard deviation data. There is a total of 24 feature data as model inputs inclinding the mean and standard deviation. The original data were reduced to three groups by using principal component analysis(PCA), which contained higher than 88% cumulative variance contribution rate. The scores of three PC with the mean and standard deviation of angular second moment together, 9 data, were used as nondestructive identification model inputs.There Soft independent modeling of class analogy model, partial squares discriminant analysis, GA-BP network model, T-S fuzzy neural network model and random forest classifier model builded in this research. Respectively in different model analyze and compare the identification ability with different sample information as inputs after different pretreatment.The main research results of this paper are as follows:(1) The identification of soybean cultivars could be feasible by the characterization of the samples with hyperspectral data. The accuracy of identification is adjusted by combining different pretreatment methods with different identification models. Average spectral data, standard deviation spectral data from and both of them together are used as T-S fuzzy neural network model or random forest classifier model inputs respectively. Their recognition results are ideal, the recognition rate of the training set is above 94%, and the test set is higher than 84%. The four identification model with image feature as inputs all hase ideal results, their recognition rate of the training set are above 93%, and the test set are higher than 96%.(2) Hyperspectral information as characterization samples, compare the identification results in the case of same model and different pretreatment to choose which pretreatment more suitable for the model in this research. To T-S fuzzy neural network model or random forest classifier model respectively with the average spectral data, standard deviation spectral data or both them as inputs, median filter smoothing is better in denoising than multiplicative scatter correction or standard normal variate.(3) The mean, standard deviation of four texture features in terms of energy, entropy, angular second moment and correlation as image feature characterization samples were extracted. Comparing the score coefficient size of PC1, the mean and standard deviation of angular second moment is the biggest and most stable with the highest reliability. The ability of angular second moment to characterize the sample is the strongest and stable. So angular second moment is better than other features to enhance the credibility of the identification results a nd to identify performance more stable.(4) When hyperspectral information as characterization samples, the identification ability of T-S fuzzy neural network model is always higher than random forest classifier model. When image feature information as characterization samples, in the four models, the best training accuracy and testing accuracy were obtained in genetic algorithms back propagation(GA-BP) network, and the results of T-S fuzzy neural network model were slightly lower than GA-BP network, while T-S fuzzy neural network model with high efficiency and high identification rate in a shorter time.