Chlorophyll Content Estimation And Nutrition Stress Diagnosis Of Gannan Navel Orange Leaves Based On Hyperspectral Technology

Hyperspectral technology has the advantages of convenience,rapid,efficiency,and no damage to plants.It is an important technical mean to achieve quantitative measurement of crop growth status.The paper takes navel orange fruit trees in Gannan as the research object,and establishes a monitoring model of chlorophyll content and nutritional stress status of navel orange fruit trees through full band,effective bands,spectral index,principal component analysis(PCA)hyperspectral data.The main contents are:Explore the modeling performance of the full band data and effective bands data under different regression algorithms;Based on the correlation coefficient of the spectral data,explore the characteristic spectrum sensitive to chlorophyll a and chlorophyll b;Using the response relationship between various spectral indexes and chlorophyll,choose a highly relevant and robust spectral index for chlorophyll to construct a chlorophyll inversion model;By extracting the spectral characteristics of various deficiency leaves,a method which rapid and non-destructive detection of the deficiency state is explored.The main conclusions and main results obtained from the research are as follows:(1)Use raw spectral(RS)data,first derivative spectrum(FDS)data and second derivative spectrum(SDS)data in the range of 400-1000 nm,using 1.67 nm and 3.41 nm two wavelength intervals,through four regression methods(Partial Least Squares Regression(PLSR),Artificial Neural Network(ANN),Ordinary Least Squares Regression(OLSR)and Stepwise Linear Regression(SLR))to predict the chlorophyll content of navel orange leaves.In modeling,the full spectrum data is used as the model input vector of PLSR and ANN,and the effective wavelength is used as the model input vector of OLSR and SLR.The results show that the data predicted performance of the 1.67 nm wavelength interval is better than 3.41 nm.The PLSR and SLR predicted models established by the full band data have better performance.The F-RS-PLSR model shows the best estimate of the chlorophyll content,training set decision coefficient(C-R2)=0.92,test set decision coefficient(V-R2)=0.96,training set root mean square error(C-RMSE)=0.05,test set root mean square error(V-RMSE)=0.19,the remaining prediction deviation of the training set C-RPD=17.00,the remaining prediction deviation of the test set V-RPD=3.63.In general,using the full-band PLSR has the best modeling effect.(2)Based on the original spectrum,the first derivative spectrum and the second derivative spectrum,using partial least squares regression(PLSR)and support vector machine(SVM)to establish the prediction model of chlorophyll a and b,respectively.The experimental results show that the predicted effect of SVM modeling on chlorophyll content is better than that of PLSR.The derivative data has a good prediction effect on chlorophyll a,V-R_~2=0.999,C-RMSE=0.004 mg/g,V-RMSE=0.004 mg/g,and the original spectral data(correlation coefficient is 0.5)obtained better for chlorophyll b.The modeling effect of R2=0.965,C-RMSE=0.003 mg/g and V-RMSE=0.003 mg/g.(3)Using RS data,FDS data and SDS data,19 published indicators that have good inversion effects in estimating the chlorophyll content of plant leaves are constructed.The predicted accuracy of chlorophyll a,chlorophyll b and chlorophyll content in navel orange leaves was compared by Adaboost(AR),PLSR,RFR,DTR and SVMR regression methods.Experimental results show that AR is the best regression method whether it uses a single spectral index or multi-spectral indices as the input vector.In addition,in predicting the chlorophyll a content(R2=0.966,RMSE=0.121)and chlorophyll content(R2=0.969,RMSE=0.144),the multi-spectral indices are superior to the single spectral index.However,using a single spectral index to predict chlorophyll b is relatively good(R2=0.931,RMSE=0.045).In general,the multi-spectral indices are suitable for the estimation of chlorophyll a and chlorophyll,and the single-spectral index is suitable for the retrieval of chlorophyll b.(4)Based on near-infrared hyperspectral data at 900-1700 nm,a method for rapid and non-destructive detection of nutritional deficiencies in Gannan navel orange leaves was explored.Using hyperspectral imaging system,250 navel orange leaf samples((magnesium deficiency(MD),boron deficiency(BD),nitrogen deficiency(ND),healthy leaf(HL),zinc deficiency(ZD))push-broom scanning inspection.Randomly select 200 pieces as the training set and the remaining50 pieces as the test set.The training set and the test set do not intersect.Developed some classifiers(Adaboost(ADA),Random Forest(RF),Decision Tree(DT),Support Vector Machine(Support Vector Machine,SVM),Multinomial Logistics Regression(MLR)and k-Nearest Neighbor(KNN)),use cross-validation method to tune classifier parameters.The research results show that the recognition rate can reach 100%by modeling through the ADA method when the data is normalized and PCA processed at the same time.