Hyperspectral Ground Object Recognition Based On Multi-scale Features

With the development of hyperspectral remote sensing technology,the acquisition of geospatial information in the current era has entered the era of high-spatial-resolution.The information contained in hyperspectral data has been continuously enriched,and the description of the feature information of hyperspectral ground objects has become more and more refined,which brings difficulties to the process of feature extraction and classification and recognition of remote sensing features.In order to make better use of the hyperspectral information of images,this dissertation deeply analyzes the unique data structure of different ground objects in hyperspectral remote sensing images.Based on the theory of statistics and machine learning,three improved models for multi-scale feature extraction are proposed.By extracting the multi-scale structure and texture information of the hyperspectral images,a more comprehensive and elaborate feature description can be obtained,and high-precision recognition of hyperspectral ground objects can be achieved.Four public hyperspectral images datasets are used to evaluate the effectiveness and generalization of the proposed method.The main work of this dissertation can be summarized as follows:(1)Aiming at the problem that traditional single-scale feature extraction methods have structural unity in expressing image features,the multi-scale analysis method is proposed for feature extraction in this dissertation,which fuses the structural distribution characteristics of various ground objects in hyperspectral images at different scales.Effectively improve the recognition accuracy and reduce the "salt and pepper" phenomenon in the process of ground object recognition.(2)From the spatial features of hyperspectral images,this dissertation proposes a ground object recognition model that integrates multi-scale grayscale and LBP features.While retaining the local gray-scale structure and texture features of the image,the model enhances the description of the structural features of real objects at different scales,and effectively improves the feature expression ability of hyperspectral remote sensing objects.(3)From the frequency domain features of hyperspectral images,this dissertation proposes a ground object recognition model that integrates multi-scale grayscale and LPQ features.The multiscale grayscale structure of the image and the Fourier transform of the LPQ feature information in the frequency domain are extracted and fused,then a more distinguishable multi-scale regional feature descriptor in the frequency domain of the image is obtained,which shows excellent performance in the feature recognition of hyperspectral images.(4)In order to further solve the limitations of the proposed multi-scale models in the spatial and frequency domains,a multi-scale feature fusion model combining of multi-scale grayscale,local binary pattern,and local phase quantization features is proposed.Through the appropriate feature fusion strategy,various multi-scale information is integrated to achieve more completed and refined feature extraction and high-precision of hyperspectral images ground object recognition,which laying the foundation for further analysis and application of hyperspectral remote sensing data in the future.