Multi-spectral Image Fusion With Projection Substitution And Low-rank Sparse Matrix Decomposition

With the rapid development of remote sensing technology, a large number of remote sens-ing images can be acquired with the aid of optical satellites such as IKONOS, QuickBird, GeoEye-1 etc. These images play an important role both in our daily life and military mis-sions such as weather forecast, disaster monitoring, military mapping and so on. Due to the limitation of sensor technology, the single-band panchromatic image acquired by those optical satellites is high in spatial resolution but poor in spectral information, while the multi-band multispectral images are characterized as rich spectral information but low in spatial resolution. However, the high spatial resolution multispectral images are much more desir-able in practical applications, because they are beneficial to image understanding and inter-pretation, by which images can be efficiently utilized and a reliable decision could be made. Benefit from data fusion, researchers propose to fuse multispectral images with panchro-matic image to acquire high spatial resolution multispectral images. The aforementioned procedure to derive high spatial resolution multispectral images is known as multispectral and panchromatic image fusion, which is also referred as pansharpening.According to the development course of pansharpening technology, highlight on the spatial distortion and spectral distortion raised in pansharpening, and lever by wavelet transform, low-rank and sparse matrix decomposition, sparse representation and image super-resolution etc., some related explorations and studies are investigated and given in this dissertation. The works in this dissertation are summarized and given as follows:1). Highlight on the spectral distortion raised in the classical principal component anal-ysis pansharpening approach, an improved version is proposed after analyzing the reason of spectral distortion. The smoothing filter-based intensity modulation technology is first employed to modulate the low-resolution spatial information in multispectral images into a high-resolution. The ATWT is further utilized to decompose the modulated spatial informa-tion and panchromatic image into low and high frequency wavelet coefficients. Different fu-sion rules are designed to fuse those two sets of low and high frequency coefficients to obtain a fused high-resolution spatial information. Experiments demonstrate that the fused spatial information is much more similar to that in original multispectral images, which indicates that the fused high-resolution spatial information is a good candidate as a high-resolution spatial information. Accordingly, performance of the classical principal component analysis approach is also efficiently improved.2). Highlight on the high redundancy and high correlation of multispectral images, con-sidering the characteristics of decorrelation and dimensionality reduction of low-rank and sparse matrix decomposition theory, the low-rank and sparse matrix decomposition theory is explored to investigate and address pansharpening problem. A low-rank and sparse matrix decomposition based pansharpening approach is proposed, and some beneficial conclusions are concluded from the studies.3). Due to the inherent limitation of component substitution based pansharpening approach-es, spectral distortion that exists in the sharpened images may be serious. After analyzing the influence of neighborhood pixels’spatial correlation, the context based decision mod-el is employed to exploit the neighborhood pixels’spatial correlation to further improve performance of the aforementioned approach. Meanwhile, according to the experimental analysis and the inherent characteristic of multispectral images, the optimal parameters for the low-rank and sparse matrix decomposition are analyzed and determined. It also demon-strates that after the neighborhood pixels’spatial correlation is considered, performance of the aforementioned approach is significantly improved.4). Lever by low-rank and sparse matrix decomposition, the function of the three compo-nents, i.e., the spatial information, the spectral information and the detail information, which compose the sharpened multispectral images are discussed. Highlight on the spatial infor-mation, the sharpness of spatial information for pansharpening is explored and investigated, and a high frequency information injection-based pansharpening approach is proposed. The study demonstrate that the sharpness of spatial information is positive to the improvement of pansharpening.5). Highlight on the derivation of detail information in classical GS1 approach, the influence of detail information for pansharpening is discussed and analyzed. Based on the assumption of scale and contrast invariant, lever by sparse representation and optical imaging model, the detail information is directly derived from the spatial features of panchromatic image. Studies demonstrate that the detail information derived in this way is much more similar to that extracted from the ideal high-resolution multispectral images, which illustrates that the directly derived detail information is much more suitable for injection. Accordingly, performance of the classical GS1 approach is improved.