Estimation Of SPAD Value And Nitrogen Content Of Maize Leaves Based On UAV Hyperspectral Image

Nitrogen nutrition status is closely related to maize yield.As the main producing area and dominant area of maize planting in China,it is of great significance to monitor the nitrogen nutrition status of maize quickly and accurately for China’s food security.In this study,spring maize in Lishu County,Jilin Province from 2019 to 2020 was taken as the research object.The hyperspectral images of canopy UAV,SPAD value of maize leaves and leaf nitrogen content(LNC)were collected at the key growth stages of maize(jointing stage,silking stage and maturity stage).The correlation between narrow band vegetation index extraction and spectral transformation and plant nutritional status was analyzed.The performance of four modeling methods of stepwise regression(SMLR),BP neural network(BPANN),random forest(RF)and XGBoost in the estimation of crop nitrogen nutrition status at jointing stage,silking stage,maturity stage and whole growth period of maize was compared.The estimation model with the highest accuracy was determined,and the contribution rate of the indicator factors in the model was calculated.The results showed that:(1)The spectral response of SPAD value and LNC of maize leaves in different spectral transformation methods was basically the same.The reflectance of the original spectrum and the first derivative(FD)transform spectrum showed an upward trend in the visible and near-infrared bands.The reflectance of the envelope removal(CR)transform spectrum and the reciprocal logarithm(Log)transform spectrum decreased in the visible and near-infrared bands.(2)RF and XGBoost models performed better than the other two regression models in the accurate estimation of leaf SPAD value and LNC in the key growth period of maize.The estimation accuracy of SPAD value and LNC based on RF model was higher than that of single growth period.The optimal estimation accuracy of maize leaf SPAD value was R~2=0.84,RMSE=6.18,RE=13.42%,and the optimal estimation accuracy of maize LNC was R~2=0.90,RMSE=0.40,RE=22.65%.(3)In the estimation of maize SPAD value,FD transform spectrum had the highest contribution rate in the construction of RF model based on the whole growth period,with a contribution rate of 27.64%,and the high contribution rate bands were mainly concentrated in the green light(550 nm)and red edge(700～780 nm).The contribution rate of narrow band vegetation index to RF model construction in the whole growth period was 14.15%,mainly including EVI,MSAVI and NDSI.The contribution rate of CR transform spectrum and Log transform spectrum to the construction of RF model in the whole growth period was small(contribution rate<10%).In the estimation of maize LNC,FD transform spectrum had the highest contribution rate in the construction of RF model based on the whole growth period,with a contribution rate of 31.91%,and the high contribution rate bands were mainly concentrated in the green band(500～600 nm)and near infrared(780～950 nm).Narrow band vegetation index,CR transform spectrum and Log transform spectrum contributed little to the construction of RF model during the whole growth period(contribution rate<10%).Therefore,it is feasible to estimate the SPAD value and LNC of maize leaves based on UAV hyperspectral images.The accuracy of the model can be improved by estimating the agronomic parameters of the whole growth period through the RF model combined with the FD transform spectrum.The results of this study can provide data basis and decision support for non-destructive,rapid and dynamic monitoring of nitrogen nutrition status of maize in black soil region of Northeast China.