Research On Raman Spectroscopy Identification Method Of Japonica Rice Seeds Based On Neural Networ

Rice has a leading position in China and is considered as one of the most crucial food crops globally.Rice variety identification as an agricultural detection technology has always been a focus of attention.China has a wide distribution of rice planting areas and numerous varieties,and different regions have different varieties suitable for planting rice seeds.However,due to inadequate market management and incomplete variety information,farmers are unable to quickly and accurately distinguish rice seeds with the naked eyes,and often encounter the problem of selecting and purchasing wrong rice seeds.Therefore,rapid identification and identification of rice seed varieties in different rice planting areas is of great significance.Currently,rice seed identification methods have changed from inefficient and difficult manual methods to high-precision and easy manipulation physicochemical methods.Recently,Raman spectroscopy has become widely available in the field of variety identification due to its fast speed,high efficiency,and convenient operation.With development of artificial intelligence,ANN has turned a research highlight in the fields of qualitative analysis and quantitative detection.Therefore,this article uses Raman spectroscopy analysis technology combined with neural network methods as a research tool to achieve efficient and accurate determination of rice seed production areas and varieties.The main contents are as follows:(1)Collection and preprocessing of spectral information.Obtain spectral information of japonica rice seeds from 30 varieties in six regions(Heilongjiang,Liaoning,Jiangsu,Henan,Shandong,and Tianjin)through Raman spectroscopy technology,and divide the data into training sets and test sets.We also studied and compared normalization(NL),Savitzky-Golay(SG)smoothing,first derivative(1-Der),second derivative(2-Der),multiplicative scatter correction(MSC),standard normal variate transformation(SNV)and adaptive iterative reweighted penalized least squares(AIRPLS)spectral preprocessing methods and their combinations have a total of 11 pretreatment approaches and the identification effect of the original spectral data in the model,from which the best preprocessing method is selected.The established models include four neural network algorithms for classification and identification:BP,RBF,PNN,and GRNN.Taking japonica rice seeds from different regions as the research object,in the BP,RBF,PNN,and GRNN neural network models,the best preprocessing methods for identification are AIRPLS+1-Der,MSC,AIRPLS,and AIRPLS,respectively,with corresponding recognition accuracy rates of 84%,86%,82.67%,and 83.33%.(2)Selection of different feature wavelength extraction methods.Due to the large amount of spectral information and the redundant data,the modeling speed is slow.Therefore,four methods are used to extract spectral data feature values,namely,clipping method,competitive adaptive reweighted sampling(CARS),successive projections algorithm(SPA),and CARS-SPA algorithm,to study and compare the identification effects of different feature band extraction methods in BP,RBF,PNN and GRNN neural network classification discriminative model,and select the best feature band extraction method corresponding to different neural network models.Taking japonica rice seeds from different regions as the experimental subject,in the BP neural network model,the CARS feature extraction method significantly shortened the modeling time and the model recognition accuracy was as high as 83.33%,with the best identification effect;In RBF and GRNN neural network models,the accuracy rate of feature extraction modeling and recognition by CARS method has been improved to 86.67% and 85.33%,respectively,with the best identification effect;In PNN neural network models,the accuracy rate of feature extraction modeling and recognition by clipping method has been improved to 83.33%,with the best identification effect.(3)Optimize different neural network analysis models.In order to achieve the better model identification effect,genetic algorithm(GA)and particle swarm optimization(PSO)are used to better the model.When taking japonica rice seeds from different regions as the research object,the BP model chooses the optimal number of hidden layer neurons,and uses GA algorithm to better the BP model.The final recognition precision of the model is 93.33%;In RBF,PNN and GRNN neural network models,PSO algorithm is used to optimize the spread value to improve the model recognition accuracy.Finally,after selecting the optimal spread value,the recognition accuracy of these three neural network models is 90%,92.67%,and 89.33%,respectively.When taking different varieties of japonica rice seeds from the same region as the research object,the data were subjected to optimal preprocessing,optimal feature wavelength extraction,and algorithm optimization.The corresponding optimal recognition and classification models for Heilongjiang,Liaoning,Jiangsu,Henan,Shandong,and Tianjin regions were RBF,RBF,BP,BP,PNN,and GRNN neural network models,with corresponding recognition accuracy rates of 90%,92%,93.33%,93.33%,95%,and 100%,respectively.This indicates that the BP neural network model is superior to other classification and recognition models for japonica rice seeds in different regions;In the classification and recognition models for different varieties of japonica rice seeds in the same region,there are corresponding optimal network recognition models for different regions.By conducting classification and recognition research on multiple regions and varieties of japonica rice seeds,research methods and reference basis have been provided for rapid and accurate identification of rice varieties.