Research On Identification And Classification Of Surface Micro-plaques Based On Hyperspectral Remote Sensing

In recent years,grasslands have been degraded to varying degrees due to the deterioration of natural conditions and the destruction of human behavior.Grassland degradation is one of the main manifestations of desertification,and rat holes,bare soil,sand spots and vegetation are specific manifestations of micro-plaques on the surface of desertified grasslands.They are also important indicators for evaluating grassland degradation.Hyperspectral remote sensing has weak information and quantitative detection.The advantages of using hyperspectral to identify and classify the surface micro-plaques are the basic research work of grassland degradation monitoring.This experiment takes ground hyperspectral remote sensing data as the research object,collects data from desertified steppe in Wulanchabu Siziwang Banner,Inner Mongolia Autonomous Region,and by machine language learning and convolutional neural network learning to find a microplaque identification and The best method for classification,so as to achieve high-precision identification and classification of surface micro-plaques.It provides a theoretical basis for the recognition and classification of desertification grassland surface micro-plaques by UAV hyperspectral remote sensing.In this paper,GaiaSky-mini hyperspectral instrument is used to collect hyperspectral imagery of desertification grasslands and rat hole quadrats.One is based on traditional machine language learning(K-means clustering algorithm,iterative self-organizing data analysis algorithm,support vector Machine algorithm)combined with dimensionality reduction methods(Principal Component Analysis,Minimal Noise Separation)to process ground hyperspectral data,using traditional machine language and dimensionality reduction analysis methods to identify and study micro-plaques on the surface of desertified grasslands.The second method is to classify and study the micro-plaques on the surface of desertified grassland by combining the method of feature band extraction(2D-PCA),artificially increased data set(WGAN)and convolutional neural network(C-LeNet5,GoogLeNet).Research shows that after the combination of the minimum noise separation method and the support vector machine algorithm,the overall recognition accuracy is 95.6%,and the kappa coefficient is 0.9342,and the classification accuracy of the surface processed micro-plaques using WGAN-GoogLeNet to identify the surface micro-plaques reaches 97.50%.It has achieved high accuracy classification,and it is accurate and practical for the identification of micro-plaques on the surface of desertified grasslands.Through machine language learning and deep convolutional neural network learning to identify and classify the surface micro-plaques of desertified grasslands,and then provide basis and methods for monitoring and quantitative inversion of grassland degradation by UAV hyperspectral remote sensing.