| With the continuous development of technology, increasingly number of observation satellites of different kinds with superior performance are becoming available. Especially with the launch of many satellites carrying remote sensing platform containing multisensors, the amount of remotely sensed data has achieved a highest level ever. But any single sensor still has certain defect and limit in performance. Thus, in order to make full use of available sensors and sensed data, and obtain more precise and valuable data under existing conditions, multisensor image fusion for remote sensing has become a hot issue.First, the development progress of remotely sensed image fusion has been reviewed, also the classical algorithms for remotely sensed image fusion have been analyzed. The object of this part of work is to mine the ideas behind different algorithms and how these ideas have been improved together with the algorithms. It is clear that researchers promote multisensor image fusion from using simple transforms for better visual effects to using complex tools for enhancing certain features in the fused data because they had realized that it is not enough to make a nice-looking image out of all input images, the ultimate demand for remotely sensed image fusion is to extract all useful information from data obtained and achieve the convert from data to knowledge.Then, this thesis analyzes the characteristics of space-borne SAR image and visible image, and formulate the whole flow of pre-processing and fusion stage. Having integrated data source chracteristics and basic ideas of image fusion, it is easier to figure out what algorithm we need for fusion of space-borne SAR and visible images.After that, NMF (Non-negative Matrix Factorization) has been introduced and compared with traditional data representation tools like PCA and VQ. Then NMF is demonstrated to be an appropriate tool for remotely sensed image fusion because its unique ability of unsupervised local learning. But when there is certain defect in input imagery, for instance, some areas in visible image are blocked by clouds, the performance of NMF may fall. WNMF (Weighted Non-negative Matrix Factorization) is introduced to deal with this situation. This thesis uses five different weighting operators to compose different WNMF algorithms.The last chapter gives results and quantitative evaluation of the preprocessing methods and fusion algorithms introduced earlier. The main point is to compare traditional fusion methods with NMF and WNMF algorithms based on different operators. Also there raises a discuss of applicable conditions for NMF and WNMF based on the results. All the results and evaluation have concretely prove that NMF and WNMF are all effective methods for fusion of space-borne SAR and visible images, and they perform extremely well at the aspect of integrating distinct information from different sensors. Applying NMF and WNMF in a flexible way could provide satisfying performance for space-borne remotely sensed image fusion in different applications. |
