Visual Sparse Representation And Deep Ridgelet Networks Based Remote Sensing Image Fusion And Classification

With the development of the satellite technology, the high-resolution remote sensing image appear constantly, and the remote sensing images have been used widely used in the field of agriculture, urban area analysis, environmental monitoring, road network extraction and so on. A variety of remote sensing image can be obtained by the different imaging equipment, such, multi-spectral image, hyperspectral image, and panchromatic image. In order to better interpret the object information, multi-spectral image and panchromatic images fusion and hyperspectral image classification are studied in this paper.In this thesis, in order to improve the quality of the fused image and the classification accuracy of the hyperspectral image, the two aspects, the fusing method based on the visual sparse representation and the hyperspectral image classification method based on the deep ridgelet networks, are studied in this paper. All works contained in the thesis are summarized as:(1) A novel fusion method is proposed in this thesis for multi-spectral and panchromatic images via Mask Dodging and non-sampled shift-invariant shearlet transform. It is shown that the proposed method can solve the impact of thin cloud on the fused image. In this method, the thin clouds can be removed by the Mask Dodging method effectively. And for the cloud removal multi-spectral, the fusion method based on the adaptive PCA transform and non-sampled shift-invariant shearlet transform is proposed to balance the spectral and spatial information. Since the clouds removal process makes the detail lost problem, an image enhancement method is designed to enhance the details of the cloud regions in the fusion process. Several experiments shown that the proposed method can maintain not only the better spatial resolution and spectral information in the fused image, but the more consistent clarity.(2) A novel fusion framework is proposed in this thesis for multi-spectral and panchromatic images via the primal sketch model and learning interpolation. Based on the primal sketch model, primal sketch map in the high-resolution panchromatic image can be obtained. Primal sketch map consists of several segments, which contain the edges and lines features of the image. And each segment of the primal sketch map has its own direction. According to the direction of the segments, a new regional segment method is designed to segment the multi-spectral and panchromatic images into structure and texture, and smooth regions. For the structure and texture regions in the multi-spectral image, combine the gradient information of the panchromatic image, structure and texture learning interpolation methods are proposed to learn the interpolation pixels. It is shown by experiments that the proposed method is effective on several different kinds of multi-spectral and panchromatic images.(3) A fusion method is present in thesis for multi-spectral and panchromatic images based on compressed super-resolution reconstruction and multi-dictionary learning. A two stage compressed super-resolution reconstruction model is proposed to solve the problem that the errors exist between the high-resolution panchromatic image and the linear weighted high-resolution multispectral image. Meanwhile, combine the sketch information of the image, a multi-dictionary is designed by compose the ridgelet dictionary, curvelet dictionary, and DCT dictionary according to region characteristics of multi-spectral and panchromatic images. This dictionary is used for both stages to improve the fusion results. It is shown that the proposed method can solve the spectral distortion problem in recent compressed sensing fusion methods, and can further the spatial information in the fused image.(4) Based on the standard autoencoders proposed by Hinton, a novel ridgelet autoencoders is present in this thesis. Based on ridgelet autoencoders, hyperspectral image classification method is proposed. Ridgelet is considered as the activation function to further improve the sparse approximate ability of the high-dimension non-linear decision function. Compared with the commonly used activation function includes sigmoid, hyperbolic tangent and rectifier functions, ridgelet has scale, location, and orientation information; meanwhile, it is compactly support. Hence the sparse approximate ability can be further improved by the proposed ridgelet autoencoders. Experiments are tested on the hyperspectral images, MNIST dataset, CIFAR10 dataset, and double helix structure, a conclusion is given that ridgelet autoencoders can improve the classification accuracy effectively.(5) According to the framework of convolutional neural networks, combined the characteristics of ridgelet, a novel frameworks of ridgelet convolutional neural networks is proposed in this thesis for hyperspectral image classification. In this method, ridgelet filter initialization method and spectral-spatial hyperspectral image classification method is proposed. Compared to the standard initialization methods, ridgelet initialization has more rapid approximation rate in approximating the spectral and spatial information. Compared with the standard feature learning methods, the proposed method utilizes the framework of convolutional neural networks, and provides an adaptive learning method for learning effective features. It is shown by experiments that the proposed method can improve the approximate rate and classification accuracy effectively.(6) A novel framework for three-dimension convolution neural networks is proposed in this thesis for hyperspectral image classification. Standard convolutional neural networks have been demonstrated the excellent performance in feature learning, but it can learn the features from the two-dimension (2-D) image only. Hyperspectral image is a pixel cube, and the space characteristics are also useful for classification to improve accuracy. Hence we extend the 2-D CNNs into 3-D for hyperspectral image classification problem. By performing 3-D convolutions, more complex space features can be exploited. Besides, 3-D ridgelet initialization filter is also studied in this paper. It is shown by experiments that the classification accuracy is improved greatly.