Retrieving Nutrient Information Of Japonica Rice Based On Unmanned Aerial Vehicle Hyperspectral Remote Sensing

At present,the application amount of chemical fertilizer in the rice production process of Northeast China increases year after year which has caused many ecological problems to some extent such as partly land degradation,environmental pollution and increase of plant diseases and insect pests.Therefore,the implement of quick and nondestructive test on its nutrient information during the rice growth process has important significance for auxiliary implementing the rice nutrition diagnosis,precision fertilization and enhancing fertilizer utilization rate,and reducing environmental pollution,etc.At present,due to the limits of spatial resolution and spectral information,the rice nutrient information inversion models based on satellite remote sensing and ground remote sensing are difficult to meet the requirements of region class rice nutrient information accurate inversion.Along with the rapid development of drone and hyperspectral technologies,the new method and technical support have been provided to solve the region class rice nutrient information accurate inversion problem.This study takes the northeast japonica rice as the main study object and carries out the rice field cultivation experiment based on the "3414" fertilizer design in the Daonan Experimental Field of Shenyang Agricultural University for two consecutive years from 2015 to 2016,Integrate the drone high spectrum remote sensing platform to obtain the high spectrum remote sensing images of rice canopy during different rice growing periods,adopt the spectral angle mapping method,maximum likelihood classification,Fisher discriminant,support vector machine method and second generation wavelet fusion algorithm to extract the characteristic information of rice and classify.The classification result shows that the classification accuracy of pure rice hyperspectral information extracted through the second generation wavelet classification based on the expectation maximization algorithm is 90.36%which is higher than the rice hyperspectral information extracted through other classification algorithms.The classification result shows that the pure rice hyperspectral information in the complex rice field environment can be comparatively ideally extracted through the second generation wavelet classification algorithm.According to the analy,sis of hyperspectral characteristics of northeast japonica rice,the hyper spectrum of rice is mainly that the variation of spectral information is determined by the pigment content in leaf within the visible light region of 400nm to 750nm,the absorbencies of rice to the blue band and red band are relatively better,the absorbency on the green band is relatively weaker than that on blue band and red band,the rice leaf spectral characteristic are mainly affected by the variation of cellular structure within the near-infrared band of 750nm to 1000nm.Within the near-infrared band,the reflectivity of rice on light is relatively strong and the absorbency is relatively weak.According to the analysis of leaf high spectral characteristics,the variation tendencies of reflectivity in the front and back of rice are consistent and there is no obvious difference in the visible light region,the reflectivity in the front of rice is slightly higher than the reflectivity of back within the near-infrared region.Comparing the fresh leaf with dry leaf,the reflectivity of fresh leaf is obviously less than the reflectivity of dry leaf on the whole,the spectral information variations caused by different rice nutrient contents are the important theoretical basis for this paper to inverse the rice nutrient information based on the hyperspectral information.Use the multispectral vegetation index and hyperspectral characteristic information and adopt the regression analysis method to build the rice nutrient information inversion models of rice chlorophyll,nitrogen,leaf area index(LAI)and biomass,etc.The model inversion results show that the determination coefficients of chlorophyll,fresh biomass,dry biomass,LAI and nitrogen inversion models based on multispectral vegetation index are 0.498,0.485,0.414,0.599 and 0.542 respectively.The determination coefficients of chlorophyll,fresh biomass,dry biomass,LAI and nitrogen inversion models based on hyperspectral vegetation index are 0.617,0.569,0.615,0.690 and 0.668 respectively.The effect of rice nutrient information model based on the hyperspectral characteristic information inversion is better than the rice nutrient information inversion model based on traditional multispectral,but the rice nutrient information inversion model built through the regression analysis method is susceptible to be affected by external factors.Through the optimized crop canopy radiative transfer mechanism model PROSAIL,this paper proposes the rice nutrient information inversion model N-PROSAIL.Comparing with the existing PROSAIL inversion model of rice nutrient information,the N-PROSAIL model can make up the weakness that the existing PROSAIL model cannot conduct the rice nitrogen inversion.On this basis,respectively use the lookup table method and numerical optimization method to conduct the rice nutrient information inversion.The determination coefficients of models that adopt the N-PROSAIL model to inverse the rice nutrient information model are 0.712 of chlorophyll,0.565 of fresh biomass,0.696 of dry biomass,0.696 of LAI and 0.709 of nitrogen respectively.The inversion effect of model is better than the inversion accuracy obtained by regression analysis.Use four machine learning algorithms of genetic neural network,Gaussian process regression,kernel ridge regression and random forest to build the rice nutrient information inversion model.Among them,this study adopts the Gaussian radial basis kernel function to conduct the coring of ridge regression algorithm and convert it into kernel ridge regression algorithm which can comparatively ideally process this kind of nonlinear problem of rice hyperspectral while lowering the model input parameter.The accuracy of rice nutrient information inversion model obtained by kernel ridge regression inversion in this study is better than other three kinds of machine learning algorithms and the determination coefficients of fresh biomass and leaf area index models are 0.723 and 0.786 respectively which are higher than the inversion models built by other methods.The study results of this paper can provide the theoretical basis and technologic support for the quick and nondestructive testing and scientific fertilization of rice nutrient information.