Research And Application Of Remote Sensing Image Fusion Method

Remote sensing image is an information carrier recording the reflective electromagnetic wave of the earth detected by the satellite sensors, which is valuable for environmental monitoring, land-cover classification, weather forecasting, etc. Many Earth observation satellites provide remote sensing images at different spatial and spectral resolution, such as multispectral images and panchromatic image. Multispectral images have high spectral resolution but low spatial resolution, while panchromatic image has high spatial resolution but low spectral resolution. In order to benefit from both spectral information of the multispectral images and spatial information of the panchromatic image, these two kinds of images are often fused to get one high-spectral-and high-spatial-resolution remote sensing image.The existing methods can be sorted into component-substitution methods, multi-resolution analysis methods and model-based methods. The most commonly used component-substitution methods such as those based on the intensity-hue-saturation (IHS) transformation and principal component analysis (PCA) often lead to spectral distortion. Multi-resolution analysis methods such as wavelet-based image fusion methods may cause spatial distortion. Furthermore, model-based methods like maximum a posteriori estimation-based fusion method requires that panchromatic image and multispectral images are highly correlative, and iterative least square estimation-based fusion method is sensitive to noise in the observation equation.Our contributions described in this thesis are shown as follows:First, we proposed a remote sensing image fusion method based on generalized IHS (GHIS) transformation and compressive sensing. Compressive sensing model from high-resolution intensity component to low-resolution intensity component and high-resolution panchromatic image is constructed. Then, the high-resolution intensity component is recovered using basis pursuit method. Finally, panchromatic image in the GIHS-based fusion method is substituted by the recovered high-resolution intensity component, and the fused multispectral images are obtained. By applying the high-resolution intensity component, the spectral distortion of the fused images is reduced greatly.Second, a remote sensing image fusion algorithm for multispectral and panchromatic images based on robust estimation is proposed. The observation model from high-resolution multispectral images to low-resolution multispectral images and high-resolution panchromatic image is constructed. Then robust estimation is introduced into the maximum a posteriori framework to get the fused images. By introducing robust estimation, the observation noise effect on the estimation results is reduced greatly. And the qualities of the fused images are improved.Finally, the classification of the unfused and fused multispectral images is processed using Support Vector Machine and Random Forest classifiers. By comparing their classification results, we come to the conclusion that high spatial resolution and good spectral character result in higher classification accuracy. It is also proved that the fusion methods proposed in the thesis are more effective on classification of remote sensing images than several other methods.