Research On Detection Method Of Soybean Nutrient Based On Multiscale Hyperspectral Imaging Technology

Soybean[Glycine max?L.?Merr.]is an important food,oil and feed crop in the world.It is also an important source of high quality vegetable oil and vegetable protein.It plays an important role in the dietary structure and national economy of China.In recent years,there are some problems in soybean production in China,such as low yield,poor quality and weak market competitiveness.As a result,soybean production in China is highly dependent on foreign countries and the contradiction between supply and demand is becoming increasingly acute.Fertilization is an important measure to improve the yield and quality of soybean.At present,there is a lack of scientific fertilization guidance in agricultural process and excessive use of fertilizer.This not only improves the production cost,but also brings a series of environmental pollution problems.Rapid and accurate determination of nutrient status in soybean growth process is an important prerequisite for scientific fertilization,therefore,it is urgently needed to develop methods forrealizing the rapid and accurate determination of nutrient status in soybean growth process.Taking soybean as the research object,the indoor hyperspectral imaging data of soybean leaves,ground hyperspectral imaging of soybean canopy and remote sensing images of low altitude simultaneously of soybean field were obtained based on hyperspectral imaging technology.The relationship between the acquired hyperspectral imaging data and soybean nutrient content was analyzed and the quantitative detection models for soybean nutrient content were established.The results would provide scientific guidance for the rational use and dynamic regulation of fertilizer during the growth of soybean plants.The main research achievements are shown as follows:?1?The hyperspectral image data of soybean leaves were obtained by using the indoor hyperspectral imaging acquisition system.The relationship between spectral reflectance of soybean leaves and the content of N,P and K were analyzed combined with laboratory chemical analysis methods.The fullwavelength variables models and characteristic variables models for the rapid and non-destructive detection of N,P and K were establishedrespectively.And the visualization of nutrient content distribution in soybean leaves was realized.In the analysis of full wavelength variable,different preprocessing methodslike Savitzky-Golay smoothing?SG?,multiplicative scatter correction?MSC?,standard variable normalization?SNV?,De-trending,first derivative?1-Der?,second derivative?2-Der?and Direct Orthogonalization Signal Correction?DOSC?were utilized to eliminate the noise.The effects of different spectral preprocessing methods on the performance of partial least squares?PLS?model were compared.The best spectral preprocessing methods for the determination of N,P and K content detection were DOSC,1-Der and SNV,and the prediction determination coefficients for thethree best models were 0.9428,0.7157,and 0.8944,respectively.Competitive adaptive reweighted sampling?CARS?,uninformative variable elimination?UVE?,successive projections algorithm?SPA?,genetic algorithm?GA?,independent component analysis?ICA?,and random frog?RF?were used to extract the characteristic wavelength.Based on the selected characteristic variables,the partial least squares regression?PLS?,least squares-support vector machine?LS-SVM?and extreme learning machine?ELM?models were established and compared.Results showed that the best model for the estimation of N content was PLS model with the characteristic variables selected by DOSC-RF,andthe corresponding prediction determination coefficients was 0.9466.The best model for the estimation of P content was PLS model with the characteristic variables selected by1-Der-GA,andthe corresponding prediction determination coefficients was 0.7465.The best model for the estimation of K content was LS-SVM model with the characteristic variables selected by SNV-UVE,andthe corresponding prediction determination coefficients was 0.9075.Based on the optimal prediction model for N,P and K and the pseudo-color image coding technology,the distribution maps of N,P and K content in soybean leaves were generated respectively.The visualization of soybean nutrient information was realized at leaf scale.?2?Soybean canopy hyperspectral imaging data was acquired using a field hyperspectral imaging acquisition system.The relationship between the spectral reflectance of soybean canopy and the contents of nitrogen,phosphorus and potassium was investigated.The models for N,P and K determination at the canopy scale were established.The canopy spectra were preprocessed by SG,MSC,SNV,De-trending,1-Der,2-Der and DOSC.The PLS modeling method was used to determine the optimal spectral preprocessing method for the determination of N,P and K content.Results showed that DOSC was the best preprocessing methods for the determination of N,P and K.The determination coefficients of the optimal preprocessing model for N,P and K were 0.9377,0.8701 and 0.8211 respectively.On the basis of spectral preprocessing,the characteristic variables were selected by CARS,UVE,SPA,GA,ICA and RF algorithms respectively.PLS,LS-SVM and ELM models were established and compared to determinate the best determination models for N,P and K content using the characteristic variables.The best model for the detection of canopy N content was UVE-LS-SVM,and the determination coefficient for prediction set is 0.9447.For the determination of canopy P content,the best model wasbuilt by SPA-LS-SVM,and the determination coefficient of this model for prediction set is 0.8775.For the determination of canopy K content,the best model was SPA-PLS,and the determination coefficient of this model for prediction set is 0.8271.According to the optimal prediction model and the pseudo-color image coding technology,the distribution maps of canopy N,P and K contents were generated respectively.?3?Correlation coefficient analysis,stepwise regression and spectral index were used to select the spectral characteristic variables and the optimal spectral indices which were closely related to the nutrient content of soybean at the region scale.Based on the selected characteristic variables and the optimal spectral indices,the multiple linear regression modelsfor the determination of N,P and K content were established and compared respectively.The optimal spectral indiceswere the basis for building the best multiple linear regression model of N,P and K content.The determination coefficients of the best model for N,P and K content were 0.9063,0.8072 and0.5632,respectively.The corresponds spectral indices of the best N content detection model were NDSI(R₅₅₂,R₅₅₅),RSI(R₅₃₇,R₅₇₃)and DSI(R₅₄₀,R₅₅₅).The corresponds spectral indices of the best P content detection model wereNDSI(R₅₄₉,R₅₇₃),RSI(R₅₄₀,R₅₇₃)and DSI(R₄₈₃,R₄₈₆).The corresponds spectral indices of the best K content detection model wereNDSI(R₆₅₇,R₆₇₂),RSI(R₆₇₂,R₆₅₄)and DSI(R₆₆₀,R₆₇₂).Using the acquiredhyperspectral image of UAV and the best prediction modelsfor N,P and K content,the spatial distribution mapsof N,P and K content weregenerated at regional scale duringthe flowering and seed filling stage.The distribution maps were consistent with the actual results acquired at the ground.The spatial information of soybean nutrient status could be reflected in theimage of UVA basically.The results provide the foundation for the rapid,dynamic and non-destructive monitoring of soybean nutrient status at regional scale.?4?The application of fractional differential in the detection of nitrogen content in soybean canopy was studied.The correlation between the preprocessed spectra and soybean canopy nitrogen content pretreated bydifferent fractional order differential?0² order differentials with the differential interval of 0.1?were analyzed using the normalized difference spectral index?NDSI?and ratio spectral index?RSI?.Results showed that the useful information in spectral data could be extracted and refined using fractional differential algorithm.In addition,the sensitivity of spectra to soybean canopy nitrogen content could be enhanced.Specifically,the positive correlation between soybean canopy nitrogen content and the band near red edge platform,and the negative correlation between soybean canopy nitrogen content and the band near green region were enhanced.Quantitative correction model of nitrogen content in soybean canopy under different order differential was established and compared with the prediction model of nitrogen content in soybean canopy established by common vegetation index.The model based on the ratio spectral index RSI_0.7(R₅₄₈,R₇₆₇)under 0.7 order differential had the best performance in estimation of soybean canopy nitrogen content with the determination coefficients of prediction of 0.8003,root mean square errors of prediction of 3.5111 and ratio of prediction to deviation of 2.2537.The results indicated that fractional differential algorithm had certainadvantages in the quantitative estimation of soybean canopy nitrogen content.