Nitrogen Nutrition Diagnosis Of Rice Based On Computer Vision And Hyperspectral Technology

Nitrogen,phosphorus and potassium are the main nutrient elements in the process of rice growth,and nitrogen is particularly important.During rice planting,the application of nitrogen has a great impact on the whole growth of rice.Nitrogen application will more or less affect the yield and quality of rice.Reasonable nitrogen application in rice can not only save nitrogen resources,protect environment,but also effectively improve rice yield.In order to explore a rapid and non-destructive diagnosis method for nitrogen nutrition in rice,the method of combining computer vision technology,hyperspectral technology and machine learning was used to carry out the research on nitrogen nutrition diagnosis in rice.In order to realize the quantitative and qualitative diagnosis of nitrogen nutrition status of rice based on computer vision technology,the super hybrid rice Liangyoupeijiu was used as the test variety,and four different nitrogen application levels were set for rice cultivation experiment.During the young panicle stage and the full heading stage of rice,the images of the top first fully expanded leaf leaves(1st leaf),the top second fully expanded leaf leaves(2nd leaf),and the top third fully expanded leaf leaves(3rd leaf)were obtained by scanning respectively,and 25 color and geometric morphological features were obtained by image processing technology.In quantitative diagnosis,25 rice features of three leaf positions were selected to predict the total nitrogen content of rice leaves by multiple linear regression modeling.The sensitive leaf positions of rice in different periods were compared and analyzed,and the total nitrogen content estimation model of sensitive leaf position of rice was established by using BP neural network and support vector machine.In qualitative diagnosis,19 color and geometric features of rice3rd leaf at the young panicle stage were selected,and the BP neural network optimized by genetic algorithm was used to establish the classification and recognition model of rice nitrogen nutrition.It is concluded that the geometric feature data measured by image processing method can effectively replace the manual measurement,and the average relative error of leaf length and width is only 0.328%and 3.404%,respectively.In the quantitative diagnosis,the 3 rd leaf at young panicle differentiation stage and the 2nd leaf at the full heading stage were more sensitive than other leaves at the same period,and the young panicle stage was the most sensitive;the total nitrogen content estimation model of rice sensitive leaves based on machine learning was slightly better than multiple linear regression model,and BP neural network was the best model,RMSE_v=0.090,MRE_v=0.034,R_v~2=0.89 about the verification set of 3rd leaves at the young panicle stage,RMSE_v=0.132,MRE_v=0.046,R_v~2=0.820 about the verification set of 3rd leaf at qualitative diagnosis.In qualitative diagnosis,the BP neural network model optimized by genetic algorithm is better than the single BP neural network model and the BP neural network model optimized by traditional genetic algorithm.The average total recognition rate of the test set samples of the established rice nitrogen nutrition classification recognition model is 99.000%.In order to realize the qualitative diagnosis of nitrogen nutrition status of rice based on hyperspectral technology,Zhongjiazao 17 was used as the test variety,and four different nitrogen application levels were set for rice cultivation experiments.The spectral reflectance of the 3rd leaf of rice at tillering stage from visible light to near infrared(350-2500 nm)was obtained by spectral analyzer.At the same time,three different parameter optimization algorithms were used to optimize the support vector machine(SVM)to establish the classification and recognition model of rice nitrogen nutrition.The parameter optimization algorithms include grid search algorithm,particle swarm optimization algorithm and genetic algorithm.The results show that the spectral reflectance curves of rice leaves under different nitrogen application levels are roughly the same,but there are certain differences in spectral reflectance of the 780-1300 nm,1400-1850 nm and 1900-2500 nm bands under different nitrogen application levels.Compared with the SVM model under default parameters,the classification and recognition effect of training set and test set of SVM with parameter optimization is better.And,the SVM model optimized by genetic algorithm has the best performance,and the average total recognition rate of training set and test set sample is 99.375%and98.750%,respectively.The results show that the application of computer vision technology and hyperspectral technology combined with machine learning methods can perform well in the diagnosis of rice nitrogen nutrition,which provides a new way for rapid rice nitrogen nutrition,and provides technical support and theoretical basis for accurate nitrogen application.