Research On Rice Variety Identification Model Based On Hyperspectral Image Technology

The quality and authenticity of rice seeds are directly related to subsequent rice growth,processing and yield.In recent years,with the rapid development of hyperspectral image technology,it has been applied to the quality inspection and variety classification of rice,which has the characteristics of scientific rapid,non-destructive and high precision.The experiment selected the hybrid rice planted in the experimental field of Anhui Agricultural University:Nonghuayou 808,Indica rice,Xuanliangyou,Gangyou 364,Xiangjingnuo,Liangyou 608 and Tianyou 998 as the research object,which is based on hyperspectral Image technology identifies and classifies rice varieties.The experiment proved that the extracted hyperspectral data of rice and the multi-characteristics of rice seed images can realize the non-destructive classification of rice varieties,and the two complement each other to some extent.The main research contents of this paper are as follows:(1)In this study,100 kinds of hyperspectral images in the 908-1735nm range of 7 varieties of rice were collected.The reflection information of rice hyperspectral image was extracted by ENVI software to represent the sample characteristics,including the reflectance spectrum mean and the reflectance spectral standard deviation,and the rice spectrum was extracted.The image seed feature information represents the sample features,and preferably five morphological feature parameters(circumference,area,squareness,circularity,elongation)and five textures in four directions of 0°,45°,90°,and 135°.Characteristic parameters(energy,entropy,contrast,local stationarity,correlation).(2)For the hyperspectral image of rice samples,black-and-white plate correction and adaptive filtering pre-processing denoising were used.The standard deviations of the extracted reflectance spectra and reflection spectra were corrected by SG,SNV and MSC.When hyperspectral data is used as model input,the PSO-PNN neural network identification model and the random forest combination classifier identification model are constructed.The influence of different hyperspectral data preprocessing methods on the accuracy of the discrimination results is compared and analyzed.(3)Seven spectral bands were selected by continuous projection algorithm for spectral data corrected by multivariate scatter.The morphological and texture features of rice spectral images under characteristic bands were extracted.The morphological features and texture features are combined into the reflectance spectrum mean and the reflectance spectral standard deviation,and BP neural network model,SVM discriminant model and CS-SVM discriminant model are constructed respectively.The effects of rice spectral information and image feature information on the accuracy of each model’s identification results were compared and analyzed.The main conclusions of this paper are as follows:(1)The rice spectral information can be used to identify rice varieties.Different spectral data preprocessing methods have different precision in the identification model.The reflectance spectrum mean,the reflectance spectral standard deviation and the fusion data of the two are used as the PSO-PNN neural network identification model and the random forest combination classifier model input.The classification results can achieve better accuracy,and the training set accuracy rate is 92.Above%,the test set accuracy rate can also reach more than 82%.Under the same input conditions of the same model,the experimental results produced by the three spectral data preprocessing methods are also different.It is found that the data processed by the MSC method is more accurate than the other two methods in the identification model selected in this experiment.SNV Transform Processing takes the second place,and SG results are the worst.At the same time,it is found that the PSO-PNN neural network classification and identification model achieves the highest accuracy in the identification of rice varieties after the multi-scattering correction of the reflectance spectrum and the standard deviation of the reflectance spectra as the model input.The training set accuracy rate is 99.05%,and the test set accuracy rate is 97.14%.(2)The combination of rice spectral information and image features can identify rice varieties.The experimental methods are used to extract multi-feature information of rice seeds in spectral images at characteristic wavelengths,and the spectral characteristics of rice are integrated to construct four classification models.The results show:among the three identification models,the classification model established by using cuckoo search algorithm to optimize SVM is more accurate than SVM.The side reflects that the optimization algorithm can effectively improve the accuracy of discriminant classification.Among the three identification models,the CS-SVM model has the highest recognition accuracy,the training set accuracy rate is 98.29%,and the test set accuracy rate is 95.43%.It is also stable and can be used for the identification of rice varieties.