Nitrogen Balance Index Prediction Of Winter Wheat Based On Canopy Hyperspectra

Nitrogen is the most important nutrient element required for the growth and development of winter wheat.Both the deficiency and excess of nitrogen will have a certain impact on the yield of winter wheat.Therefore,it is very necessary to monitor the nitrogen status of winter wheat in real time.In this research,winter wheat in Guanzhong area in 2016,2018 and 2020 was taken as the research object.The collection of canopy hyperspectral information and the determination of nitrogen balance index were carried out during each key growth period of winter wheat,and the original canopy hyperspectrum was conventionally transformed,including smoothing,first derivative,continuum removal and logarithmic transformation,and on the basis of conventional transformation spectrum,carry out continuous wavelet transformation and discrete wavelet transformation.Based on the above transformed spectra,the nitrogen balance index machine learning model was constructed from the two directions of characteristic bands and spectral indices.Estimating the nitrogen balance index through the model provides certain theoretical and technical support for regional winter wheat nitrogen monitoring and fertilization guidance.The main conclusions of this study were as follows:(1)As the growth stage of winter wheat progressed,the nitrogen balance index and spectral reflectance changed significantly.Among them,the spectral reflectance showed a downward trend with the continuous use of nitrogen from the jointing stage to the booting stage.From flowering stage to filling stage,Nitrogen balance index tends to be stable after fluctuating.The correlation analysis between the transformation spectrum and nitrogen balance index of each growth stage showed that,except for the filling stage,the correlation coefficients between the transformation spectrum and nitrogen balance index of other growth stages were concentrated in(-0.8,0.8).The correlation analysis between the spectral index and the nitrogen balance index extracted based on the transformed spectrum at each growth stage showed that the MTCI,DCNI and VOG spectral indexes under all transformed spectra at the jointing stage had a very significant correlation with the nitrogen balance index,and the spectral index of the continuum removed during the filling stage,the correlation coefficient between DCNI spectral index and nitrogen balance index was the largest(0.96).(2)The nitrogen balance index estimation model was constructed based on the characteristic bands and spectral indices(any two-band combination of spectral index and vegetation index)under the conventional transformed spectrum.The results showed that in the nitrogen balance index estimation model established based on the characteristic bands,the logarithmic transformed spectral partial least squares regression model at the jointing stage had an excellent performance.To estimate the sample ability,the coefficient of determination of the verification set R~2was 0.76,and the root mean square error RMSE was2.13;the first-order derivative spectral support vector machine regression model of the flowering stage had an excellent ability to estimate the sample,RPD was 2.35;The first-order derivative spectral random forest regression model of the filling period had a rough estimate of the sample capacity.In the nitrogen balance index estimation model based on the spectral index(any two-band combination of spectral index and vegetation index),the partial least squares of the continuum removal spectrum at jointing stage,the first derivative spectrum at booting stage,the smoothing spectrum at the flowering stage,the first derivative spectrum at the filling stage and the continuum removal spectrum at all growth stage.The multiplicative regression model had an excellent ability to estimate samples,with RPDs of2.46,2.08,2.67,3.41,and 2.27;the coefficient of determination R~2of the verification set of the continuum removal spectrum random forest regression model at the filling stage was0.93,RMSE was 2.03,and RPD was 3.46.(3)In the continuous wavelet transform,the db5 wavelet mother function performed better in each growth stage.Spectra with better correlation with nitrogen balance index after continuous wavelet transform at each growth stage were continuum removal spectra except flowering stage which was the first order derivative spectrum.The partial least squares regression model verification set coefficient of determination R~2based on the S9 scale characteristic parameters in the flowering stage was 0.94,the RMSE was 1.65,and the RPD was 3.14;the support vector machine regression model based on the S10 scale characteristic parameters in the filling stage had an excellent ability to estimate samples(RPD was 2.33).(4)In discrete wavelet transform,compared with coif5 and sym8 wavelet mother functions,the db5 wavelet mother function had a stronger compression ability(the number of coefficients on the S10 scale accounted for 0.95%),and it can better highlight the spectral profile information(after the approximate coefficient was reconstructed,it was the same as the original spectral signal was highly similar).The continuous projection algorithm extracted the least feature parameters(3)at the jointing stage S10 scale,and the most feature parameters(25)at the filling stage S1 scale.The verification set coefficient of determination R~2of the partial least squares regression model based on the S1 scale characteristic parameters in the flowering stage was 0.96,the RMSE was 1.32,and the RPD was 3.94;the support vector machine regression model based on the S2 scale characteristic parameters in the filling stage had excellent estimation samples ability(RPD was 2.20);the random forest regression model based on S1 scale characteristic parameters for the whole growth stage had a rough estimation of sample ability(RPD was 1.7).(5)Comparing the accuracy of the nitrogen balance index estimation model of different methods,the results showed that:the spectral index model(PLS-CR model)performed better in the jointing stage(RPD was 2.46),followed by the characteristic band model(PLS-LOG model);The spectral index model(SVR-FD model)in the heading stage performed better(R~2was 0.78,RMSE was 1.93),followed by the discrete wavelet transform spectral characteristic parameter model(PLS-S8 model);the discrete wavelet transform spectral model in the flowering stage(PLS-S1 model)performed better(RPD was 3.94),followed by the characteristic band model(PLS-LOG model);the discrete wavelet transform spectral model(PLS-S6 model)performed better during the filling stage,followed by the spectral index model(RFR-CR model);the spectral index model(RFR-FD model)performed better in the all growth period(RPD was 2.63),followed by the continuous wavelet transform spectral model(PLS-S1 model,RPD was 2.52).