Research On Monitoring And Analysis Of Soybean Stress State Based On Hyperspectral Technology

In recent years,there is a growing attention in the agricultural field for crop growth monitoring by spectral technology of electromagnetic radiation.Among them,hyperspectral technology is an important part of this monitoring technology,and has been widely used due to its high resolution and non-destructive monitoring advantages.Aiming at the problems of inaccurate and inefficient monitoring of crop growth stress status,this paper takes soybean as the research object,hyperspectral technology was employed as the main method to address the three common stresses of health,water and nitrogen stress,and disease stress during soybean growth.Explore the response characteristics of soybean hyperspectral and physiological information under different stress states.The methods of monitoring and identifying different stress states of soybeans were studied.The main research contents and results are as follows:1.Prediction of physiological information under unstressed(healthy)stateHyperspectral data and net photosynthetic rate of chlorophyll content of soybean leaves were collected.Built seven original and seven first derivative spectral indices respectively,and pairwise combinations of spectral reflectance in the band of 325-1075nm were performed by the correlation matrix method.Calculate the correlation coefficient with the two kinds of physiological information,use the maximum correlation coefficient as the extraction standards to select the optimal wavelength combination and spectral indices.The optimal spectral indices were divided into three groups as model input variables.Compared with seven commonly used spectral indices to evaluate predictive ability of the optimal spectral indices.Three methods of partial least squares regression PLSR,least squares support vector machine regression LS-SVMR and LSSSO regression LASSOR were used for modeling.A soybean physiological information prediction model was established based on the optimal spectral indices and common spectral indices.The model was calibrated and verified by greenhouse data and field measurement data,respectively.The results showed that the predictive ability of physiological information based on the optimal spectral indices was better than the commonly used spectral indices.The LS-SVMR method performed the strongest predictive ability of physiological information,while the LASSOR method showed the weakest predictive ability.The model determination coefficients R_c² and R_p²,root mean square error RMSEC and RMSEP of calibration set and verification set were used as evaluation indicators.The optimal spectral indices based on the combination of original and first derivative showed the highest prediction accuracy for chlorophyll content,and the R_c²=0.9410,R_p²=0.9496,RMSEC=1.4164,RMSEP=1.3759 of the chlorophyll content optimal model.The prediction accuracy of net photosynthetic rate based on optimal spectral indices calculated by the first derivative spectral was the highest,and the R_c²=0.8661,R_p²=0.7978,RMSEC=1.3920,RMSEP=1.5411 of the optimal model.2.Methods for monitoring water and nitrogen stressIn this paper,the response characteristics of physiological information and hyperspectral under water and nitrogen stress were analyzed,the sensitive spectral indices were selected by the correlation coefficient method,then the monitoring research of water and nitrogen stress was conducted respectively based on the original hyperspectral and sensitive spectral indices.The results showed that based on the analysis of the original hyperspectrum,the hyperspectral curves with the lowest reflectance in the 500-700nm band and the highest reflectance in the 760-900nm band represented no water and nitrogen stress,the gradual increase of reflectance in the 500-700nm band indicated the increase of stress degree.Five sensitive spectral indices were selected from the 15 spectral indices,which were normalized differential vegetation index(NDVI),green normalized differential vegetation index(GNDVI),modified red-edge normalized vegetation index(m NDVI₇₀₅),chlorophyll index(LCI)and ratio vegetation index(RVI).Their correlation coefficients with two kinds of physiological information were both greater than 0.8.The maximum values of the five sensitive spectral indices represent no water and nitrogen stress,and the values decrease gradually with the increase of water and nitrogen stress.3.Physiological information inversion under water and nitrogen stressIn this paper,three smoothing methods,SG convolution smoothing,fast fourier transform FFT and wavelet transform smoothing WT were used,a total of 12 preprocessing methods included multiple scattering correction MSC,variable standardized SNV,first derivative FD and second derivative SD four single methods were employed,partial least squares PLS and principal component regression PCR methods were used for modeling.The correlation analysis method was used to select the characteristic bands,and the characteristic bands were taken as the input variables of the model.The soybean physiological information inversion model was established by combining with different pretreatment methods and modeling methods,and was compared and analyzed with the model established by full-band hyperspectrum.The model correlation coefficient r_c and r_p,root mean square error RMSEC and RMSEP were used as evaluation indexes to determine the optimal physiological information inversion model.The results showed that the characteristic wavelengths had better physiological information retrieval ability than the full wavelengths.The MSC+FD+S-G+PLS method was the best method to establish the chlorophyll content inversion model,r_c=0.9606,r_p=0.9702,RMSEC=0.9600,RMSEP=1.0300 with the optimal chlorophyll content model.The SNV+SD+S-G+PLS method was the best method to establish the net photosynthetic rate inversion model,r_c=0.9927,r_p=0.9708,RMSEC=0.3260,RMSEP=0.6120 with the optimal net photosynthetic rate model.4.Disease classification and identificationTwo common diseases of soybean,frogeye leaf spot disease and bacterial blight disease,were studied.The original hyperspectral was processed and analyzed by three pretreatment methods,MSC+SG,SNV+SG and FD+SG respectively,and was processed by two characteristic wavelengths selected methods,competitive adaptive reweighted sampling CARS and successive projections algorithm SPA,respectively.Based on fisher discriminant analysis and LS-SVM modeling methods,the recognition accuracy of single category and overall category of the confusion matrix of the prediction set was taken as the evaluation criteria to confirm the optimal disease classification recognition model.The results showed that all the models can achieve better disease classification and recognition.No matter which modeling method was adopted,SNV+SG+SPA method can achieve single category and overall classification accuracy of 100%.Furthermore,based on different characteristic wavelengths selected methods,the hyperspectral preprocessed by SNV+SG showed a strong classification ability,the SNV+SG-LS-SVM model established by combining it with the LS-SVM method can achieve disease recognition with a classification accuracy of100%.5.Disease classification diagnosisThe severity diagnosis of frogeye leaf spot disease and bacterial blight disease was studied.Hyperspectral and disease image data were collected,by referring to the classification method of disease class in Agriculture Industry Standard of China,the disease class was determined by the feature extraction of disease image based on computer vision technology.Based on this grade,the disease class was diagnosed by hyperspectral technology further.Two methods were used to reduce the dimension of spectral data,extraction of characteristic wavelengths based on CARS algorithm and extraction of optimal spectral indices based on PCA respectively.The LS-SVM class diagnosis model of disease was established based on the reduced spectral data.Taking the classification accuracy as the evaluation standard,the feasibility of the data dimensionality reduction method was evaluated and the optimal disease classification model was selected.The results showed that after dimensionality reduction by CARS algorithm,20 and 52 characteristic wavelengths were extracted for frogeye leaf spot and bacterial blight respectively,and the number of wavelengths decreased by 94.67%and86.13%,respectively.The ability of disease classification of spectral data after dimensionality reduction was better than that of full-wavelength spectrum.For frogeye leaf spot,the models built based on the data sets DS2,DS8 and DS9 performed the highest overall classification ability,with classification accuracies of 94.5%,97.3%and 97.3%,respectively.Compared with the accuracy of full-band spectrum(93.6%),the accuracy increased by 0.9%,3.7%and3.7%,respectively.For bacterial blight,the models based on DS2,DS11 and DS16 had the best classification performance,which can achieve the overall classification accuracy of96.8%,95.8%and 94.7%,respectively.Taking the full-wavelength spectrum(91.6%)as the benchmark,the accuracy increased by 5.2%,4.2%and 3.7%,respectively.