Fusion Of Multispectral And Panchromatic Iamge Based On Multiscale Transform

It is desirable to have the images with both high spectral and spatial resolutionfor photo analysis and improving the results of a number of applications like landcover classification, change monitoring, and object detection. The process ofpansharpening aims at increasing the spatial resolution of multispectral (MS) imagesas well as preserving their spectral signature by the help of the spatial information ofhigh-resolution panchromatic (Pan) images. Recently, a family of multiresolutiontransforms named X-lets based fusion methods have been proposed to overcome thelimitation of the wavelet from the perspective of the expression of directionalinformation and intrinsic geometrical structures. The aim of this paper is to assesscurvelet, contourlet, NSCT and shearlet based fusion methods and to investigate thatwhich one is better suited for pansharpening.Firstly, the definition of pansharpening is discussed. The theory of X-letstransforms and ARSIS concept is reviewed. Four multiscale models based on X-letstransforms are constructed to fulfill the fusion task, while three general MRA-basedimage fusion algorithms are presented. Two of them are developed under the ARSISconcept. Another one can be labeled as a hybrid model incorporating both PCA andX-lets.In addition, the index Q4is generalized to Q8for8-band WorldView2(WV2)image using the theory of octonions. Both theoretical analysis on extensibility ofindex Q and experimental results prove the proposed index Q8inherits the subtlediscrimination capability of spectral distortion and of inaccuracies in spatialenhancement from Q4, successfully.In the last, a thorough comparison among curvelet, contourlet, nonsubsampledcontourlet transform (NSCT) and shearlet transform based fusion methods isperformed in three different injection model: ARSIS global model, ARSIS localmodel and hybrid model. Visual and objective comparisons are presented onQuickBird (QB) and WV2image data, as well as the robust assessment. The resultsshows that NSCT and shearlet transform outperform the others. Concerning thecomputation complexity, shearlet transform is recommended for pansharpening.