Image Fusion and Segmentation Algorithms for Remote Sensing Applications

Abstract

In remote sensing, the fusion of high spatial resolution Panchromatic (PAN) images with low spatial resolution Multi-Spectral (MS) images (after spatial co-registration) is very useful technique to get high spatial resolution data over wide swath at low cost i.e. without actually going for a costly satellite sensor of both high spatial resolution and MS in nature. The challenge of fusion techniques is to preserve as much as possible both the spatial and spectral correlations to the maximum extent, so that the fused data can be used for quantitative radiometric analysis, such as segmentation or classi cation. The image fusion is also useful in many applications like discrimination of ground covers, ne features on earth's surface, identifying textures and boundary of object. In most satellite sensors mismatches between the coverage of the PAN and MS bands causes color distortion. Many image fusion algorithms are available in literature to reduce the color distortion. Image segmentation is crucial to object oriented remote sensing imagery analysis, to locate objects, to identify boundaries in images and for the change detection studies. If the radiometry of the pixel is changed then it may be considered being part of other segment. So the quality of the fusion algorithm a ects the accuracy of the segmentation. In this report, the standard image fusion algorithms available in literature to reduce the color distortion are explained. Two di erent image fusion algorithms are proposed. The rst fusion algorithm is based on di erence between MS image and it's mean. It also involves histogram matching of each MS image to that of PAN image. Histogram matching of each fused image to that of MS image keeps the mean and range of each image similar to that of MS image. The range of the fused image changes because of the di erence between the images and addition of PAN image. Hence the range correction is required to be done on the fused images to get the same range as in the MS image. The second algorithm is based on injection of high frequencies into the MS images. The amount of injection of high frequency, factor K, depends upon Normalized Di erence Vegetation Index (NDVI). The NDVI indicates the type of surfaces like vegetation, soil, water and urban etc. The relationship between how much high frequency should be injected into the MS image, factor K, and NDVI is found for the IKONOS-2 sensor. The combined relationship is also found for the IKONOS-2,WorldView-2 and QuickBird-2. Di erent fusion algorithms available in literature and two proposed algorithms are compared by visually as well as recently proposed quality assessment parameters. The quality of the fusion algorithm may a ect the accuracy of the segmentation. The images are classi ed by the unsupervised K-Means, and supervised Maximum Likelihood classi ers. It is evident from the experimental results that the Mean Di erence, High Frequency Injection, Adaptive Component Substitution algorithms reduce color distortion, increases spatial resolution, improves classi cation accuracy. Hence the proposed algorithms are well suited for the image fusion and segmentation.