The Spatial Resolution Improvement Of Optical Remote Sensing Image With Regularization Methods

In recent years, optical remote sensing images are on their way to the higher spatial, spectral and temporal resolutions. Spatial resolution represents the target size corresponding to one pixel in the image, and is a critical index for evaluating the quality of remote sensing images. For the area of earth observation, high spatial resolution can provide more rich and detailed details, which can greatly improve the application potential.Due to the physical limitations, after several decades of rapid development of hardware, improving the resolution of remote sensing images by producing new hardware has approaching its bottleneck, In the other way, with the fast development of digital image processing technique, it has become an important way to enhance the spatial resolution of remote sensing image with algorithms. Because of more bands and large data volume of remote sensing images, it is not satisfied to directly apply the methods for nature images to remote sensing images, which would result in low efficiency, poor adaptivity and spectral distortion. To solve these problems, based on regularization methods and exploiting the spatial and spectral properties of remote sensing images, we develop the total variation and sparse prior, and establish more suitable regularization models for image denoising, single image super-resolution and image fusion, respectively. In addition, fast algorithms have been designed to solve the models, which can effectively and efficiently improve the spatial resolution of remote sensing images, and the main contents of this study are listed as below:(1) A combined spatial and spectral weighted hyperspectral denoising method is proposed. For hyperspectral images, the noise type and noise intensities have an obvious variation in both spatial and spectral dimensions. To avoid the situation of remaining noise and missing details when using the traditional methods, we propose to first estimate the relative noise intensities, and then assign each pixel the corresponding weight, after which the combined spatial and spectral models are built with Alternating Direction Method of Multipliers (ADMM) method to solve the model. The experiments have shown that, the proposed combined spatial and spectral weighted hyperspectral denoising method can effectively remove the noise and preserve the details of hyperspectral images.(2) A single hyperspectral image super-resolution method based on subspace constraint is proposed. Traditional methods often need more observation images and it often costs much time for the large data volume. To solve this problem, we investigate the spatial and spectral subspace of hyperspectral images, and then establish the single hyperspectral image super-resolution model under subspace constraint, and finally the algorithm based on ADMM is proposed to solve the model. The experiments have shown that, the proposed single hyperspectral image super-resolution method based on subspace constraint can better recover the details and protect the spectral signatures at a relatively fast speed.(3) A practical compressed sensing based image fusion method is proposed. Taking the images to perform image fusion as observations and the fused results as the original signal, we introduce the breakthrough technique of compressed sensing into image fusion. By developing a practical method to build the dictionaries, the sparse vector can be resolved under sparse constraint and the fused image can then be reconstructed by multiplying the dictionary and the sparse vector. The experiments have shown that, the proposed practical compressed sensing based image fusion method can better combine the spatial and spectral information of remote sensing images.(4) A two-step sparse coding based image fusion method is proposed. For the practical compressed sensing based image fusion method, it still has weak points for the linear assumptions in the observation model and the burden of collecting training samples and performing dictionary training. To solve these problems, the coupled sparse representation theory is introduced, and a two-step sparse coding method which featured by the injection of the structure similarity prior of different bands into the sparse coding stage is proposed, which can tackle the unstable problem by using the traditional sparse coding methods when the vector to be coded has a low dimension and a weak structure. The experiments have shown that, the proposed two-step sparse coding based image fusion method can further improve the results with a much better efficiency, and is applicable for practical applications.