Study On Hyperspectral Estimation Of Canopy Nitrogen Content Of Rice In Cold Region

Nitrogen is an indispensable nutrient for rice growth.Precise nitrogen management and dynamic regulation are the key measures to ensure good quality and high yield of rice.Using hyperspectral technique to analyze the relationship between spectral parameters and nitrogen content is an effective way to realize dynamic and non-destructive monitoring of rice nitrogen status.As special soil background,climate condition and growth period of rice in cold region,it is necessary to find adaptable models to estimate its nitrogen content.Therefore,in this study,Daohuaxiang No.2 was used as experimental object to study the estimation of canopy nitrogen content of rice in cold region by using hyperspectral technique.The experiment was carried out in Wuchang,Heilongjiang Province in 2016.According to single factor design method,experimental field was designed with 4 nitrogen gradient treatments and 8 replications,a total of 32 plots.Firstly,hyperspectral images of rice canopy at tillering stage,jointing stage and heading stage were collected using SOC710VP portable imaging spectrometer,and the corresponding canopy leaves were cut off.Secondly,canopy hyperspectral reflectance was extracted using ENVI5.0 software and nitrogen content was measured by AA3 flow analyzer.Thirdly,5 spectral pre-processing methods,namely savitzky-golay smoothing(SG),multiplicative scatter correction(MSC),standard normal variate(SNV),first derivative(FD)and second derivative(SD),were applied in 11 strategies,namely SG,MSC,SNV,FD,SD,SG-FD,SG-SD,MSC-FD,MSC-SD,SNV-FD and SNV-SD.Original spectral reflectance of tillering stage,jointing stage and heading stage were denoised and filtered using these 11 strategies,and partial least squares regression(PLSR)estimation models were established between nitrogen content and whole wavelengths after pre-processing in order to select the best pre-processing method.On this basis,3 characteristic wavelength selection methods,namely successive projections algorithm(SPA),uninformative variable elimination(UVE)and competitive adaptive reweighted sampling(CARS)were used to select the characteristic wavelengths of tillering stage,jointing stage,heading stage and whole growth stage in 4 strategies,namely SPA,UVE,UVE-SPA and CARS.Finally,3 modeling methods,namely partial least squares regression(PLSR),radical basis function neural network(RBFNN)and extreme learning machine(ELM)were used to establish the estimation models of tillering stage,jointing stage,heading stage and whole growth stage,and the best characteristic wavelength selection method and modeling method were selected according to determination coefficient(R~2)and root mean square error(RMSE)of the models.This study presented a multi-algorithm combination model SG-FD-CARS-ELM,which was suitable for estimating canopy nitrogen content of rice in cold region.Among them,SG could filter out high-frequency noise that was existed in original spectral,and FD could eliminate other noise interference,amplify effective information and enhance spectral characteristics.CARS could eliminate redundant information and select the key characteristic wavelengths,which selected 11,10,10 and 12 wavelengths at tillering stage,jointing stage,heading stage and whole growth stage,respectively.The performance of CARS-ELM models was the best at each growth stage,R_C~2 and R_P~2 were 0.972 and 0.956 at tillering stage,0.968 and 0.952 at jointing stage,0.946 and 0.935 at heading stage,0.908 and 0.898 at whole growth stage,respectively.The results could provide a geographical reference for rapid detection of rice nitrogen content and provide technical support for guiding precise fertilization management.