Estimation Of Physiological Indexes Of Nitrogen In Wheat With Different Quality Types By Hyperspectral Remote Sensing Model

In order to realize the directional cultivation and accurate management of wheat with different quality types,field test data of wheat with 4 nitrogen levels and 2 quality types from 2018 to 2020 were set as the basis to study the correlation between canopy hyperspectral and wheat canopy hyperspectral and chlorophyll(SPAD),Leaf nitrogen content(LNC),Glutamine synthetase(GS),Grain protein content(GPC)and other physiological indicators of nitrogen,using the First derivative(FD),Second Derivative(SD),Multiplicative Scatter Correction(MSC)and Standard Normal Variable(SNV)Original spectrum(OS)was processed,and then the optimal spectral data were determined by Partial least component regression(PLSR).Successive Projection Algorithm(SPA),Competitive Adaptive Re-weighted Samplings(CARS),Successive Projection Algorithm + Competitive Adaptive Re-weighted Samplings + Successive Projection Algorithm(SPA+ CARS).Finally,the sensitive bands selected by different algorithms were modeled by partial maximum,PLSR,BP neural network(BPNN)and Support vector machines(SVM)respectively.Comparing the estimation effect of the models,the hyperspectral remote sensing estimation models for each physiological index of nitrogen were selected.The main research results are as follows:(1)The canopy OS response curves of different nitrogen physiological indexes of wheat were studied.In the visible light range(400-600nm),canopy spectral reflectance decreased with the increase of SPAD or LNC,while in the near infrared range(800-1000nm),the spectral reflectance decreased with the increase of SPAD or LNC.In the range of visible light(400-600nm),the spectral reflectance of canopy of strong gluten wheat was lower than that of weak gluten wheat.(2)The leaf SPAD hyperspectral remote sensing estimation model was constructed,and OS was processed by four algorithms respectively.Then,PLSR modeling was combined with FD.The modeling effect was the best.R2 was 0.93,the Root mean square error(RMSE)was 1.6,the R2 of the prediction set was 0.51,RMSE was 3.9,and the Relative Prediction Deviation(RPD)was 1.31.Compared with the 9 models,FD-CARS-BPNN hyperspectral remote sensing model had the best SPAD estimation effect on wheat leaves.The R2 of the modeling set was 0.87,RMSE was 2.22,and the R2 of the prediction set was 0.83,RMSE was 2.05,and RPD was 2.35.The input variables of the model were reduced by 553.The results showed that the hyperspectral remote sensing estimation model of wheat leaf SPAD could be constructed by using the pretreatment and sensitive bands,thus providing a new method for the hyperspectral remote sensing estimation of wheat leaf SPAD.(3)The hyperspectral remote sensing estimation model of wheat Leaf nitrogen content(LNC)was constructed,and OS was processed by four algorithms respectively.The model effect of preprocessed spectral data combined with PLSR was basically consistent with that of OS combined with PLSR.Therefore,Then using OS as the analysis data,OS-PLSR modeling set R2 was 0.93,RMSE 0.16,modeling set R2 was 0.73,RMSE 0.29,and RPD was 1.85.Compared with the 9 models,the OSCARS-PLSR hyperspectral remote sensing model had the best estimation effect on wheat LNC,and the R2 of the model set was 0.93,RMSE was 0.16,the R2 of the prediction set was 0.83,RMSE was 0.2,RPD was 2.2,and the input variables of the model decreased by 583.The results show that the hyperspectral remote sensing estimation model of wheat LNC can be constructed by using preprocessing and sensitive bands,thus providing a new method for hyperspectral remote sensing estimation of wheat LNC.(4)The hyperspectral remote sensing estimation model of Glutamine synthetase(GS)in wheat grains was constructed.OS was treated with FD,SD and MSC,and the modeling effects of three pretreatments combined with PLSR,BPNN and SVM were compared.Differential processing(FD,SD)combined with different models is ideal,so it is possible to use differential processing(FD,SD)to eliminate the noise that exists in OS.Compared with different models,FD combined with BPNN and PLSR has an ideal modeling effect.Among them,FD-PLSR has the best estimation effect,with R2 of its modeling set 0.94,RMSE 0.025,R2 of its prediction set 0.76,RMSE 0.034,and RPD 2.31.The results showed that the hyperspectral estimation model of GS in wheat grains could be constructed by using preprocessing combined with different algorithms,so that the hyperspectral remote sensing could be used to estimate the GS activity of wheat grains in the field.The nondestructive and rapid detection of GS in wheat grains could provide technical guidance.(5)The hyperspectral estimation model of Grain protein content(GPC)was established,and it was found that SPAD had the best prediction effect on GPC at the anthesis stage,and the determination coefficient R2 was 0.63.The 4 algorithms were processed with OS respectively,and then combined with PLSR modeling respectively.FD combined with PLSR had the best effect.The modeling set R2 of FD-PLSR was 0.85,RMSE was 1.37,the prediction set R2 was 0.58,RMSE was 2.13,and RPD was 1.31.Comparing the effects of the 9 models,FD-SPA-CARS-BPNN model had the best ability to predict leaf SPAD at flowering stage,with R2 of 0.86,RMSE of 1.36,R2 of 0.86,RMSE of 0.92 and RPD of 2.78 in the model set.Therefore,FD SPACARS-BPNN can be used to indirectly predict GPC with chlorophyll as the intermediate variable,and its modeling set R2 is 0.52,RMSE is 1.2,and its prediction set R2 is 0.84,RMSE is 0.51,and RPD is 2.38.Results show that can take advantage of the pretreatment and sensitive wavelengths to build the flowering period of wheat leaf SPAD high spectral estimation model,and then use this model to estimate indirectly GPC,the mature period of wheat with high spectral estimate GPC provides a new path,thus,in the early stage of the wheat growth in different type wheat field directional cultivation quality and accurate management to provide technical guidance.