Effects Of Image Fusion Algorithms And Spatial Resolution Ratio On Fused Result

With the developing of remote sensing techniques and the using of different satellites, different series of multi-spatial, multi-spectral and multi-temporal resolution images cover the same area are sent back to the earth. Generally, for the images from the same sensor, when the spectral resolution is high, the spatial resolution is corresponding low and vice versa. To utilize these images with different characteristics efficiently, image fusion is desired to combine the low spatial resolution multi-spectral image with high spatial resolution panchromatic image in order to achieve a multi-spectral image with high spatial resolution. The fusion process overcomes the shortage of single source for interpretation.Since the early 1980s, developing image fusion techniques has been drawn more and more attention under the requirements of different application. In recent years, many image fusion algorithms have been proposed. The ideal image fusion method is to preserve the good spatial and spectral information. Meanwhile, the ideal image fusion should be applied to image classification well. As many different spatial resolution images are available, making full use of these images with different spatial resolution and choosing appropriate spatial resolution ratio are important issues to be considered. The main contribibutions are listed as follows:(1) As many measures for image fusion quality evaluation are available, multivariate statistical analysis is applied to these measures to avoid redundancy of different measures. Firstly, factor analysis is employed to the correlation matrix of measures to extract factors. Then hierarchical cluster analysis is applied to the extracted factors to obtain typical measure for each factor. Finally, a total score are obtained based on the typical measures to measure the different aspects of fused images. Experimental results show that a total of 27 measures are classified into five categories and the typical measure is selected for each category. The representative measures for assessing image fusion quality are CE (Cross Entropy) for difference-based, WSNR (Weighted Signal to Noise Ratio) for noise-based, RRIQA (Reduced Reference Image Quality Assessment) for similarity-based, QILV (Quality Index based on Local Variance) for information-clarity-based and QABf (Objective Image Fusion Performance Measure) for overall-based. Moreover, a total score is proposed based on these five measures to evaluate the different aspects of fused images.(2) To study the effect of image fusion algorithms on fused results, a number of 14 fusion algorithms are performed on the QB panchromatic and QB multi-spectral images. These fusion algorithms include the ones based on algebraic operation, based on color transform, based on high-pass filtering, based on pyramid decomposition, based on wavelet transform and the ones proposed in recent years. Quantitative evaluation measures are employed to compare the different fusion results. It is found that the evaluation rankings of different fused results on the five independent measures are different. In general, the IHS transform, region-based fusion and the Brovey transform fused images obtain high quality. Moreover, the region-based fused result obtains highest quality on the total score and classification accuracy. It is also found that the rankings of different fused results on the total score are very close to that of classification accuracy of unsupervised ISODATA. It is demonstrated that the total score could be used to compare the classification accuracy of fused images.(3) Image fusion are performed on the images from Quickbird, SPOT, ETM+sensors, thus 6 combinations of panchromatic and multi-spectral images are obtained in the principle of the spatial resolution of panchromatic image is higher than that of multi-spectral image. It is found that the quality of fused result is related with the quality of original image. The high quality of original image produces high quality fused result; when the original high spatial resolution is the same, the larger spatial resolution ratio will cause the fused result worse. Meanwhile, the different spatial resolutions are obtained by resampling techniques to study the effect of ratios on fused results. These combinations are Quickbird panchromatic image fused with downsampled Quickbird multi-spectral image, SPOT multi-spectral image fused with downsampled Quickbird panchromatic image, and ETM+multi-spectral image fused with downsampled Quickbird panchromatic image, respectively. When the ratio from different sensors is large, the downsampling will reduce the ratio. It is found that when the ratio decreases, the quality of fused image decreases first and then increases. When the ratio is larger than 10, decreased to the tenth of the original ratio will obtains higher result than decreased to the other ratios.(4) Non-subsampled contourlet transform overcomes the shortage of block effects of discrete wavelet transform and preserve the spectral information of multi-spectral image well. In addition, the mean shift segmentation is applicably to high spatial resolution image. An improved region-based image fusion is proposed. This algorithm is based on non-subsampled contourlet transform and mean shift segmentation, making use of entropy as region property and taking structural similarity as weights to fuse the coefficients of sub-images. As region-based fused image having advantage on image classification, an improved region-based image fusion algorithm is proposed. Experimental results show that the proposed fusion algorithm performs better than the other five region-based fused results in total score and classification accuracy.