Study On The Classification Of Grassland Species Based On UAV Hyperspectral Remote Sensing

China's grassland area is one of the largest in the world,with 86.667 million hectares in Inner Mongolia,including 68 million hectares of effective natural grassland,accounting for 27%of the total grassland area in our country.In recent years,grassland degradation has become the primary ecological problem in country.Since the reform and opening up,the animal husbandry in Inner Mongolia has developed rapidly,but the contradiction between grassland and livestock has become increasingly prominent,and the overall ecological environment of grassland has been deteriorating.Grassland degradation is not only manifested in the decrease of grass per unit area,but also in the deterioration of all kinds of fine grass and the change of vegetation community structure.In this study,the Desertification Grassland in Inner Mongolia is taken as the research object,and the specific experimental site is the experimental area of Sizi wang Banner in Inner Mongolia.In order to get accurate grassland degradation information,two sets of hyperspectral acquisition systems,ground and air,were used.Compared with the satellite remote sensing,the hyperspectral data can ensure the space resolution of the data,and basically realize the recognition,classification a nd regional monitoring of grassland grass species.Because there is a lot of redundant information in hyperspectral data,PCA method is used to recombine the hyperspectral data on the ground and in the air,which effectively reduces the amount of data and the complexity of processing.By PCA band method,each band of hyperspectral data is regarded as a one-dimensional vector to form an uncorrelated n-dimensional matrix.By calculating the eigenvalues and eigenvectors of the covariance matrix,several bands with the largest contribution rate are obtained and recombined.Because there is a lot of redundant information in hyperspectral data,PCA band method is used to recombine the hyperspectral data on the ground and in the air,which effectively reduces the amount of data and the complexity of processing.By PCA band election method,each band of hyperspectral data is regarded as a one-dimensional vector to form an uncorrelated n-dimensional matrix.By calculating the eigenvalues and eigenvectors of the covariance matrix,several bands with the largest contribution rate are obtained and recombined.After data preprocessing,the ground and air data are input into convolution neural network.The ground data is extracted by convolution,batch normalization,excitation function and pooled down sampling to obtain the maximum receptive field.Finally,the weight coeffi cient is adjusted by back propagation to output better classification results from the full connection layer.In this study,deep layer network and shallow layer network are used for comparative experiments.The experimental results show that vgg-16 convolution neural network has better classification performance for the dominant species of desertification grassland.In this study,the input model of vegetation data in different periods is also verified,and the conclusion that the classification effect of vegetation in seed stage is the best is obtained.The air data is input into bisenet bi-directional segmentation convolution neural network after drawing the perceptual region,deriving the location information of the perceptual region,enhancing the data,and then the data is extracted from the image's space information,edge detail features,and the maximum perceptual field through the space branch network and the context branch network,respectively.The extracted features are integrated through the feature fusion module,and the output is weighted Finally,the classification results are output by the full connection layer through continuous iteration.For the aerial semantic segmentation network,the image can be visualized to show the classification results,which more intuitively shows the classification performance of convolution neural network for the dominant species of desertification grassland.In this study,the hyperspectral data of typical Desertification Grassland in Inner Mongolia were obtained by airborne dual-purpose hyperspectral spectrometer.Based on PCA and convolution neural network,two grass species classification models on the ground and in the air were constructed.The comparative experiments show that the two classification models have great advantages in the classification of grassland species in Inner Mongolia.