Study On Fusion Methods Of Multisource Optical Remote Sensing Images

The progress and rapid development of remote sensing science and technology leadsto an enrichment phenomenon of optical remote sensing images. Image fusion is aneffective way to improve the utilization rate of such great deal of remote sensing data.Established on the design of fusion methods and spectral distortion followed by, this paperattempts to do the research from three aspects, say the fusion of panchromatic andmultispectral images, multispectral and hyperspectral images, and multitemporal images.Two tasks are completed: first, in view of disadvantages of the existing fusion methods,some improvements and new ideas are put forward, and the optimization of multisourcedata fusion are realized; second, the choice of evaluation indexes among different fusionproblems were accomplished with both space quality and spectral quality taken intoconsideration. The main content and outcome of this thesis are as follows:For panchromatic and multispectral remote sensing image fusion, choosing ZY-3satellite data as an example, the traditional IHS method is improved on the simulation ofthe I component and the injection method of spatial details. The fused images areevaluated by comparing spectral curves of typical ground features and utilizing a qualitywith no reference (QNR) index. Results show that, the improved method has a spectralfidelity of89.94%, and its QNR value reaches up to0.75. Thus, the proposed methodperforms well with no significant spectral and spatial distortion.For multispectral and hyperspectral remote sensing image fusion, a new methodbased on spectral mixture model is proposed, in which spectral and spatial information arerepresented with spectral and abundance matrix of selected endmembers rather than lowand high frequency information of images. To vetify the effectiveness of this method,SPOT and Hyperion image fusion is chosed to do the experiment, and a butterworthlow-pass filter is used to revise the fusion result. Among the evaluation indexes areentropy, peak signal to noise ratio (PSNR) and spectral angle(SA). The computed datashows that, the PSNR is larger than50, and the cosine of SA calculated by every pixel islocated in [0.6717,0.9996]. Therefore, it demonstrates the proposed method an advantageon both spectral fidelity and spatial enhancement abilities. For multitemporal remote sensing image fusion, the method based on one-pair imagelearning is introduced to fuse MODIS and TM images. In this method, completedictionary training and sparse representation theory are applied to calculate the transitionimage and a high-pass filter is then used to predict the high resolution image. Simulationimages are used to analyze influence factors, and ERGAS, SSIM and NDVI scatter plotare adopted to evaluate the fusion results with real images as reference. Results show that,the quality of fusion images can be improved by reducing the registration error and ratioof spatial resolution between high and low resolution images.