Rice Yield Model Research Based On Vegetation Index Of Hyperspectral Remote Sensing Data

Rice is the main food crops in China,which is in the national and regional grain policy and economic development plan of the important basis.Therefore,the rice yield accurate prediction is of great significance to our country food security.Remote sensing data have updated timely,comprehensive,objective and accurate advantages,it is widely used to make real-time monitoring,large area of rice growing fast yield estimation is possible.The development of unmanned aerial vehicle(UAV)for remote sensing technology provides a new means of implementation,In recent years,the use of UAV's hyper spectral or multi-spectral remote sensing yield estimation caused a boom,before which is mainly composed of remote sensing technology based on remote sensing data and 3S technology in rice area of monitoring and yield estimation research has made some achievements.But the complex terrain and quality of remote sensing data,the influence of such factors as its large area of rice yield estimation has certain limitations,yield estimation accuracy is generally not high.Ac cording to the spectral data from 2015 to 2016,this paper studies a testing ground for rice production in Shenyang agricultural university,Liaoning province,using different methods of vegetation index and yield model and get reliable accuracy and stability,and the future research of reference and guidance for the development of rice yield estimation.This paper use the dual distance variable analysis method for a single production of NDVI and each period as the sum of RMB and the correlation analysis,screening better NDVI factor to build a yuan,principal component analysis and multivariate linear model,the result is comparatively ideal.By comparing the model parameters,it is concluded that the combination of June and August composite model is better than single day composite and composite model of the meadow,months fitting degree is high,it's R2 is 0.805,F value is 9.280,significant at 0.008.The relative error and root mean square error were 5.06%and 324 kg/hm2,It has the high accuracy.Principal component analysis model R2 is 0.806,F value is 20.789,the significance of 0,the result more accurate.Then using the PRI to establish regression model,the result is higher than the fit of the canopy NDVI,so yield estimation to select data is better than canopy leaf.It's R2 is 0.875,F value is 14.965,significant at 0.011,proven its relative error and the root mean square error were 3.89%and 402 kg/hm2,which has high precision.Finally combines two index was used to model the output,respectively,using the multivariate linear regression,principal component analysis and neural network algorithm to yield estimation.For multiple linear regression model obtained the R2 is 897,F value is10.701,significant at 0.021,this model is better than single use NDVI or PRI income model fitting degree is high;For principal component analysis model,it's R2is 922,F value is 19.567,significant at 0.003;Algorithm for neural network model,the fitting degree of 0.959,was the highest of all the fit of the prediction model,and can be seen from the image of the sample points are really close to the predicted curve,the prediction accuracy is higher than the previous model,the relative error and root mean square error were 1.31%and 66 kg/hm2,so the BP neural network is the most ideal method of forecasting rice yield.