Estimating Leaf Nitrogen Content In Wheat Under Different Vegetation Coverage Conditions Using Hyperspectral Techniques

Real-time and non-destructive monitoring of crop nitrogen(N)status is essential for precision N management in wheat.Hyperspectral remote sensing provides an effective method to obtain crop nutrition information,but canopy reflectance is affected by multiple factors such as canopy structure and soil background.A range of vegetation coverage conditions were created in the field experiments with three N levels and two planting densities over four growing seasons.The comprehensive influences of soil and canopy structure on canopy reflectance are taken into consideration,and then the analysis was made on the quantitative relationships between canopy reflectance and leaf nitrogen content in wheat.This paper explores the most effective methods to reduce the impact of soil background and canopy structure on reflectance and then obtain the optimum spectral parameters in order to establish the quantitative models for estimating leaf N content under different vegetation coverage conditions in wheat.And the results can contribute to offer technical support for non-destructive monitoring and precision N management of wheat growth.Firstly,we use the spectral-index method to detect leaf nitrogen content in wheat with canopy hyperspectrum under different soil backgrounds.Two approaches were applied in this study to reduce the interference of soil background.One is to modify the SR by introducing a parameter for reducing the soil effect and other one is to determine the portion of soil signature at pixel level through linear spectral mixture analysis(LSMA).Additionally,the relationship(R2)between SR(R471,R504)with LNC improved from 0.66 to 0.71 using the LSMA approach and increasing by 0.05.The modified SR(mSR=(R471+θ)/(R504+θ))showed even more significant improvement in the correlation with LNC(R2=0.78)and increasing by 0.07.The result of independent test(2014,n=243)showed that P-R2 and RRMSE were 0.74、14.77%,respectively;Therefore,modified SR was further used for evaluation of observation positions.When the height was 1 m above the canopy,the strongest correlation(R2=0.78)with the mSR was observed for backward 60°.When the sensor zenith angle was fixed at 0°,the strongest correlation(R2=0.79)was observed when the sensor was 0.5 m above the canopy.The results demonstrate that the modified SR and the LSMA could improve the performance of LNC estimation under different vegetation coverage conditions and the former performed better than the latter,and the prediction model has higher accuracy and stability.The optimal observation position was either 0° combined with 0.5 m or backward 60° combined with 1 m above the canopy.Secondly,we take advantage of approaches of hyperspectral characteristic parameters to detect wheat LNC with canopy reflectance under different vegetation coverage conditions,including continuum-removed(CR)method、red edge parameters and continuous wavelet transform(CWT)algorithm.Based on CR method,R2 and RMSE of wheat LNC to NMAD were 0.47.0.74,respectively;and results of model validation showed that P-R2 and RRMSE were 0.52、24.83%,respectively.It shows that accuracy and stability of model calibration and validation are poor when portion of soil background was large.Based on red edge parameters,R2 and RMSE of wheat LNC to DD(A680-718,A700-724)were 0.59、0.55,respectively;According to the former research,we introduce a revised band so that improved red edge area parameters were obtained.mDD(A680-718,A700-724)which predicted wheat LNC the best with S-R2 and RMSE of 0.65、0.60,respectively.P-R2 and RRMSE of model validation are 0.65、8.09%,respectively.It shows the mDD(A680-718,A700-724)can estimate wheat LNC more stable and accurate under different vegetation coverages.Through CWT algorithm,wavelet coefficients are obtained under particular scale and band in 350-1350nm;and then correlations of wavelet coefficients to LNC in wheat are calculated.The results show that wavelet coefficients of bior3.3(W516,S8)+ sym5(W432,S5)is the best wavelet features of all.The R2 of wavelet coefficients of bior3.3(W516,S8)+ sym5(W432,S5)is 0.75;while P-R2 and RRMSE were 0.74、17.51%,respectively.To sum up,the accuracy and stability of model calibration and validation based on CWT algorithm are the most stable and accurate under different vegetation coverage conditions.Last but not least,we take advantage of chemometric methods to detect wheat LNC with canopy reflectance under different vegetation coverage conditions,and the algorithms include stepwise multiple linear regression(SMLR)、principal component regression(PCR)and partial least square regression(PLS).The results show that the model based on PLS predicted wheat LNC the best with S-R2 and RMSEc of 0.80、0.39,respectively;And then we tested the accuracy of model with independent test by Rcv2、RMSEcv、Rv2 and RMSEp.The results show that Rcv2 and RMSEcv of LNC model were 0.77 and 0.41,respectively;while the Rv2 was 0.76 and RMSEp was 0.36.In addition to,we compare the performances of three chemometric methods for predicting wheat LNC under low、medium and high coverage.The results show that the accuracy and stability of model calibration and validation based on PLS algorithm are the most stable and accurate under different vegetation coverage conditions.