High Oil Soybean Recognition Method Based On Hyperspectral Technology

Soybean is rich in protein and oil,and is an important grain and oil crop in China.High oil soybeans can produce higher quality soybean oil and have a higher utilization rate of soybean raw materials during the production process.Therefore,rapid identification of high oil soybeans not only helps breeding experts cultivate more high-quality soybean varieties,but also has important significance for oil processing enterprises to save costs.For the rapid identification of high oil soybeans,this study took 58 soybean varieties from Northeast China as the research object,combined with spectral analysis technology,machine learning technology,and deep learning technology to analyze and study the spectral data of soybean seeds,in order to achieve nondestructive testing of high oil soybeans.The specific content of the study is as follows:(1)Data collection and processing.Firstly,the Soxhlet extraction method was used to measure the oil content of 58 soybean varieties,and the soybean samples were divided into high oil and low oil according to national standards.Secondly,a hyperspectral imaging system was used to collect hyperspectral images of 5800 samples from 58 soybean varieties in the 400-1000 nm wavelength range,and ENVI 5.3 software was used to extract hyperspectral reflectance data of soybean samples.This paper preprocess hyperspectral data using multiple scattering correction(MSC),standard normal variate(SNV),and Savitzky-Golay smoothing algorithm(SG),respectively;The successive projection algorithm(SPA)and competitive adaptive reweighting sampling algorithm(CARS)are used to extract feature bands from preprocessed hyperspectral data.(2)Combinatorial optimization model was established to identify high oil soybean.Analyze six sets of feature band spectral data extracted by SPA and CARS feature band extraction methods,and establish models based on support vector machine(SVM),K-nearest neighbor(KNN),and partial least squares discriminant algorithm(PLS-DA),respectively.Through the combination of different pretreatment methods,different feature band extraction methods and different modeling methods,18 combinatorial optimization models were finally obtained.The classification accuracy of each combinatorial optimization model was compared and analyzed from three aspects of pretreatment methods,feature band extraction methods and modeling methods.According to the classification results,it was concluded that the model established by using MSC-SPA to process hyperspectral data was the best,Among the 18 combination models,the MSC-SPA-SVM combination model has the best classification performance,with a test set accuracy of 94.5%.(3)A high oil soybean identification model based on 1DCNN was established and improved using multi-scale convolution and attention mechanism.A 1DCNN model composed of feature extraction modules was constructed.In response to the shortcomings of this model,firstly,a feature extraction module composed of multi-scale convolutional kernels was proposed,and a one-dimensional convolutional neural network model based on multi-scale features(MS1DCNN)was constructed;secondly,improved the shallow feature extraction layer of the 1DCNN model and established an attention mechanism based 1DCNN model(1DCNN＿AM);finally,the 1DCNN model based on multi-scale and attention mechanism(MS1DCNN＿AM)was constructed by combining the proposed feature extraction module with an improved shallow feature extraction layer.By comparing the classification performance of four models,the results show that: The test set accuracy of the MS1DCNN＿AM model reaches96.8%,with the best classification performance.In addition,by comparing the classification accuracy of the MS1DCNN＿AM model and MSC-SPA-SVM model,it was ultimately concluded that the MS1DCNN＿AM model is the optimal detection model.This study not only provided a fast detection method for oil processing enterprises,but also provided theoretical support for developing a detection device for quickly identifying high oil soybeans.