Classification Of Winter Wheat Varieties Based On UAV Hyperspectral Remote Sensing

As one of the main food crops in China,winter wheat has rich varieties.There are a lot differences among different varieties,which mainly reflected in the different requirements of soil fertility(nutrients)and water(climate).The influence of external environment on different growth cycles are also more diverse.Facing the current situation of wheat planting regionalization and decentralization in China.How to effectively distinguish and identify wheat varieties and realize real-time wheat category extraction is an important subject to realize production optimization and accurate management.The development of UAVs technology provides an effective means for real-time collection of wheat growth information.UAVs equipped with hyperspectral camera can extract wheat images of different growth cycles in a large area in real time,with relatively improved spatial resolution due to near-ground photography.This thesis set up a test filed in winter wheat research base in Wuzhi County,Jiaozuo City.Hyperspectral image data of 6 winter-wheat varieties in different growth cycles are obtained with UHD185 hyperspectral camera and Dajiang M600 UAV.Through the experiments of various dimensionality reduction methods,the classification preprocessing method is optimized.The fine classification of wheat varieties is preliminarily realized by using machine learning algorithm.The main work and conclusions are as follows:(1)The integrated system of UAV and UHD185 camera was adopted to select appropriate flight parameters and formulate corresponding camera Settings and aerial photography scheme to obtain hyperspectral images of winter wheat in the experimental field;Hyperspectral fusion was performed on the obtained original images,and the hyperspectral sub-bands were extracted and merged based on Photo Scan and IDL programs.Finally,the geographical information registration of control points is carried out to complete the processing of uav hyperspectral data.It was proved that the correlation coefficient between wheat spectrum and measured canopy spectrum was98%,which effectively retained the original spectral information.(2)Principal component analysis transform(PCA),independent component transform(ICA)and minimum noise separation transform(MNF)are used to reduce the dimension of UHD185 data.Peak density clustering(EFDPC)and neighborhood grouping(FNGBS)were used for band subdivision.The maximum likelihood method,random forest and U-NET neural network model were used to classify the dimensionality reduction results.The results showed that FNGBS＿U-NET and MNF＿RF combined classification results were the best.The highest classification accuracy and Kappa coefficient were 88.85%,87.30% and 0.8658,0.8514,respectively.The research results of this thesis have preliminarily solved the problem of weak feature classification of winter wheat,laid a foundation for further related in-depth research in the future,and provided important technical basis for the accurate management of winter wheat and other crops.42 figures,17 tables and 65 references.