Fusion of Panchromatic and Multispectral Images by Genetic Algorithms

Pan-sharpened MS is a fusion product in which the multispectral (MS) bands are spatially enhanced by the higher-resolution panchromatic (Pan) image. Most effective algo- rithms for pan-sharpening are based on multiresolution analysis (MRA), e.g., wavelets, Laplacian pyramids, wavelet frames, or curvelets. MRA approaches present one main critical point: filtering operations may produce ringing artifacts when high frequency details are extracted from the panchromatic image. In this paper, a pan-sharpening algorithm for 4-band MS data is proposed, which is not based on MRA, but it applies a Generalized Intensity-Hue-Saturation (GIHS) transformation to the MS bands. A genetic algorithm is adopted to define the injection model which establishes how the missing highpass information is extracted from the Pan image. The fitness function of the genetic algorithm which provides the algorithm parameters driving the fusion process is based on a quality index specifically designed for quality assessment of 4-band MS images. Both visual and objective comparisons with advanced fusion methods are presented on QuickBird image data. I. INTRODUCTION Multispectral (MS) observations from spaceborne imaging sensors exhibit ground resolutions that may be inadequate to specific identification tasks, especially when urban areas are concerned. Data merge methods, based on injecting spatial details taken from a panchromatic image (Pan) into resampled versions of the MS data, have demonstrated superior performances. In the last years, multiresolution analysis (MRA), based on wavelets, Laplacian pyramids, wavelet frames, curvelets, etc., has been applied to produce effective tools to help carry out data fusion/merge tasks. However, MRA approaches present one main critical point: filtering operations may produce ringing artifacts, e.g. when high frequency details are extracted from the panchromatic image. This problem does not decrease significantly any global quality index, but it may locally reduce the visual quality of the fused product in a considerable way. To avoid this problem, we propose a pan-sharpening algorithm which is not based on MRA, but it applies a Generalized Intensity-Hue-Saturation (GIHS) transformation to the MS bands and makes use of optimal parameters which are computed by a genetic algorithm. Similarly to other pan-sharpening methods based on the injection of spatial details, the proposed algorithm assumes an injection model. To overcome instability and data-dependent results which are typical of space-varying models, we propose a simple injection model in which the coefficients that equalize the Pan image before detail injection into the MS image are derived globally - one for each band - from coarser scales, similarly to previous schemes such as SDM (1), CBD (2) and RWM (3) techniques, but not a-priori defined on image statistics, e.g., variance, mean, correlation coefficient, etc. The coefficients are computed by a genetic algorithm (GA) together with the weights which define the generalized intensity of the original MS bands. The genetic algorithm adopts the Q4 quality index defined in (4) as the fitness function to be maximized for optimal fusion parameter estimation.