Hyperspectral And Multispectral Remote Sensing Study On Yield Estimation Of Rice

Rice is the most important crop in China and it is of great significance to keep a clear knowledge of the rice planting area and yield.It plays an important role of the food security in the world.With the development and application of the hyperspectral remote sensing and UAV remote sensing,the higher rate and accuracy requirement of rice product estimation has been presented.In this paper,we use both the hyperspectral data on the canopy scale of rice and the data obtained from the low altitude UAV to study the yield inversion.Through the observation of the rice in Meichuan town of Wuxue experimental area,we get the ground hyperspectral data and UAV multi-spectral data.Then we extract the vegetation index.endmember and texture by applying the stepwise linear regression,BP neural network and random forest algorithm to estimate the product of rice.And we analyzed the accuracy of different models developed by corresponding method.The main research is as follows.(1)According to the characteristics of hyperspectral data and UAV multi-spectral images we extract two kinds of hyperspectral vegetation indices,four kinds of UAV multi-spectral vegetation indices,five kinds of endmember,four kinds of texture features and other factors based on the sensitivity analysis and correlation analysis.The result shows that the highest correlation coefficient between the surface hyperspectral vegetation index and the rice yield are booting stage DVI(431,665)and milk stage DVI(892,1087),the correlation coefficients are 0.77 and 0.76.(2)With the feature factor of screening,we use stepwise linear regression,BP neural network and random forest algorithm to build models with the method of "stay one method" cross validation.The results showed that when the booting stage DVI and the milky stage DVI fo hyperspectral data and the DV were used as the independent variables,the values of R2 in models were 0.633,0.575 and 0.440,and the corresponding RMSE were 24.546 kg/mu,27.710 kg/mu and 29.704 kg/mu when the input variables were unmanned aerial plant vegetation index,end abundance and texture features,whether using linear regression or BP neural network or random forest algorithm,vegetation index performance were better than end abundance and the texture characteristics,again showed the strong vegetation index,the best models using the three kinds of research methods all see the vegetation index as an independent variable,R2 were 0.615,0.599 and 0.405,RMSE were 25.131 kg/mu,26.377 kg/mu and 30.628 kg/mu.(3)The results showed that the multi-platform integrated estimating model obtained by stepwise linear regression is more suitable for the optimal model of terrestrial or unmanned aerial vehicle(UAV)platform,which is obtained by using the principal component analysis method.The average R2 can be 0.661 and the average RMSE can be 23.568 kg/mu when we use the cross-validation method for modeling.