Pan-Sharpening Of Remote Sensing Images With Fusion Framework And Sparse Representation

With the development of remote sensing technology, remote sensing images have been widely used in various fields, such as weather forecast, environmental monitoring, earthquake monitoring, marine monitoring, military mapping, and so on. Optical remote sensing satellites, such as QuickBird, IKONOS and GeoEye-1, can provide observed images about the surface of the Earth. The observed images include low resolution multispectral images and high resolution panchromatic images. The multispectral image is composed of various bands, which has rich spectral information and low spatial resolution. The panchromatic image, which has high spatial resolution, only has one band. High resolution multispectral images are required in practical applications. However, due to the limitation of remote sensing sensor technology, the satellites cannot provide high resolution multispectral images. Fusion of multispectral and panchromatic images, which is also called as pan-sharpening of multispectral image, can effectively overcome this problem. The high resolution multispectral image can be obtained by fusing the multispectral image and the panchromatic image. Fusion of multispectral and panchromatic images is an important branch of remote sensing image fusion, which deserves further research. The main topic of the dissertation is the fusion of multispectral and panchromatic images. Focusing on the spectral distortion and spatial distortion of multispectral and panchromatic image fusion, five fusion methods, which are based on the known fusion frameworks, have been proposed by using the evolutionary algorithm, compressed sensing, sparse representation, dictionary learning and image restoration. The proposed methods are tested on remote sensing data, which are provided by the QuickBird, IKONOS and GeoEye-1 satellites.The main contributions of this dissertation are summarized as follows:1. Component substitution fusion method is an important framework of multispectral and panchromatic image fusion. For the spectral distortion of the component substitution fusion framework, an adaptive component substitution method with particle swarm optimization algorithm is proposed. The synthetic intensity component plays an important role in the component substitution fusion framework.The adaptive weights are computed by maximizing the objective function. The objective function can measure the spatial similarity between the low-scale intensity component and the low resolution panchromatic image. Correlation coefficient, mean structural similarity index and mutual information are used as the objective functions, respectively. The proposed method is tested on QuickBird and IKONOS data. Visual analysis and quality results demonstrate that the proposed method has superior performance.2. Intensity component and injection model can affect the performance of component substitution fusion framework. The ideal intensity component produces better spatial detail image, which can effectively improve the performance of the component substitution fusion framework. However, due to the spatial dissimilarities between the multispectral image and the panchromatic image, the spatial detail image has information which does not belong to the multispectral image. Hence, spatial distortion and spectral distortion are produced. To solve the above problem, an improved component substitution method has been proposed, which is based on locally linear embedding and wavelet fusion. Two high resolution images are designed to replace the panchromatic image in the component substitution framework, which can be obtained by the superresolution technique with locally linear embedding and the wavelet fusion technique. Using the "choose-max" rule, a fused spatial detail image can be synthesized by using two spatial detail images. A novel component substitution fusion method is proposed by using the synthetic spatial detail image and suitable injection model. Experimental results demonstrate that the proposed method has better performance in fusing the QuickBird and GeoEye-1 images, which can effectively preserve the spectral and spatial information.3. The basic idea of image formation model-based fusion framework is that multispectral and panchromatic image fusion can be converted to image restoration problem. Combining with the image formation models of the low resolution multispectral image and high resolution panchromatic image, a fusion method based on sparse representation and local autoregressive model is proposed. Sparsity prior and local autoregressive model are used to exactly recover the high resolution multispectral image. Sparse representation model supposes that the high resolution multispectral image is composed of a low frequency component and a high frequency component, in which the low frequency component and the high frequency component can be sparsely represented by a spectral dictionary and a spatial detail dictionary, respectively. The spectral dictionary and the spatial detail dictionary can be trained by the K-SVD algorithm. Local autoregressive model is used to improve the local spatial quality of the high resolution multispectral image. The QuickBird and GeoEye-1 data is applied to test the proposed method. Compared with the traditional fusion methods, the simulated and real experimental results demonstrate that the proposed method has better performance.4. Spectral correlation among all the bands of multispectral image is an important characteristic. Inspired by distributed compressed sensing theory, a new pan-sharpening method based on distributed compressed sensing is proposed, which considers the correlation characteristic among the multispectral bands and the sparsity characteristic of each multispectral band as prior information. The joint sparsity model assumes that each high resolution multispectral band is composed of a sparse common component and a sparse innovation component. The common component can be sparsely represented by a dictionary, which is learned from the panchromatic image patch sets. Each innovation component can be sparsely represented by a hybrid dictionary, which consists of four multiscale transform bases and the learned dictionary. The proposed method is tested on simulated and real data, which is provided by the QuickBird and IKONOS satellites. Compared with the traditional fusion methods, the proposed method has better fusion performance.5. For the local spatial dissimilarities and contrast inversion between the multispectral image and the panchromatic image, a dual spatial regularized model-based pan-sharpening method is proposed. First, the degraded model reflecting the relationship between the low resolution multispectral image and the high resolution multispectral image is used as data-fitting term to keep spectral fidelity. Second, two spatial regularization terms are used to enhance the spatial quality of the high resolution multispectral image. The global spatial similarity regularization supposes that the high pass component of each high resolution multispectral band has the similar spatial structure with the high pass component of the panchromatic image, which can effectively reduce the spectral distortion caused by the contrast inversion. Moreover, nonlocal self-similarity characteristic of each high resolution multispectral band is considered as another regularizer, which can effectively improve the local spatial quality of the high resolution multispectral image and reduce the spatial distortion caused by the local spatial dissimilarities. The proposed method is formed by combining the data-fitting term and two spatial regularization terms. Compared with the traditional fusion methods, the proposed method has better performance in preserving the spectral and spatial information.