The Research Of Remote Sensing Image Fusion Methods Based On General Fusion Framework

Fusion of high-spatial resolution panchromatic and low-spatial resolution multispectral image is the main methodology to overcome the shortcoming of senors. Fusion results provide much more information, which is necessary for many remote sensing applications, such as land cover, land use, vegetation study, environment monitor and urban study. All applications that require information of two or more images will benefit from image fusion. Remote sensing image fusion has been a hotspot in remote sensing for its importance.How to reduce the distortion of spatial and spectral information is the key problem in remote sensing image fusion. Focused on the problem, the content of this dissertation includes two parts, the research on fusion methods based on the general component substitution fusion framework (GCOS) and the general image fusion (GIF) framework, and the research on application-oriented quality evaluation of fusion products. The main work and results are listed as follows:1. Spatial information distortion problem appeared when the ratio of higher resolution to lower resolution is very low. To overcome this shortcoming, an improved fusion method based on GIF and MTF (modulate transfer function) filter, MTF-GIF, is proposed in this dissertation. A two dimensional low-pass filter is designed by measuring the MTF of imaging system from the Pan data. The filter is used to evaluate the low-resolution spatial information which is necessary for the GIF-based fusion. The validity of the MTF-GIF method on overcoming spatial information distortion problem is approved by experiments of fusion "Beijing-1" microsatellite images. A better tradeoff between keeping spatial information and spectral information can be obtained by this MTF-based fusion method than common methods.2. Spectral information distortion is the main problem when the COS (component substitution) fusion methods are adopted. The distorted information is commonly represented by saturation of images. Based on the GCOS framework, an improved fusion method, ZSD (the zero saturation distortion fusion) is developed in this dissertation. Three important parameters are adjusted to get zero saturation distortion in this method. Experiments improved that the saturation values of fusion result consist with that of initial MSI.3. An approximating linear relationship between panchromatic and multispectral images is proved in this dissertation based on remote sensing imaging theory and the optical image formation model. Mathematical deductions show that this correlativity is related to sensors and reflectance of objects. Based on the conclusion, a GIF fusion method with classification (WC) using least square (LS) algorithm, LS-GIF-WC, is proposed. Low-resolution Pan data is estimated by the LS algorithm in LS-GIF-WC. Experiments showed that the performance of LS-GIF-WC method is stable when the initial multispectral images have different bands. Better performance of keeping spatial and spectral information is obtained when using LS-GIF-WC fusion method.4. The normalized difference vegetation index (NDVI) and vegetation coverage are introduced as quantitative indices to evaluate the quality of fusion results. Fusion results are also applied for image classification. The classification accuracy is introduced as an index to evaluate performance of fusion methods. Experiment and mathematical deduction show that, the GIF-based method works well for vegetation applications, and high accuracy can be obtained when the ARSIS-based (Amelioration de la Resolution Spatiale par Injection de Structures) fusion results are applied to image classifications.