Researches On The Methods Of Unmixing And Band Selection For Hyperspectral Remote Sensing Images

The hyperspectral sensors gather images in hundreds of spectral bands, providing huge information with nano-scale spectrum resolution. However, due to the limitation of spatial resolution, the observation of a single pixel usually consists of more than one material, causing it to be a "mixed pixel". This phenomenon affects the measurement and analysis of ground surface. In order to precisely explore the spectra information from hyperspectral data, it is essential to decompose the mixed pixels into a collection of substances’spectra (endmembers) and their corresponding proportions (abundances). This process is called hyperspectral unmixing, which has become an active area for remote sensing applications in recent years. This thesis focuses on the task of hyperspectral unmixing, proposing different methods from both statistical and geometrical perspectives. Besides, for the dimensional problem of hyperspectral data, we have also made research to reduce the redundance of data. A kind of band selection approach is developed by transforming the hyperspectral data into complex networks. The main innovations of our research can be described into four aspects as follows:1. A new constrained independent component analysis approach is presented. We overcome the independent assumption of independent component analysis (ICA) by designing new objective function which accords with the condition of hyperspectral imagery. Different physical constraints are introduced into the proposed method, making it to be essentially applicable for the analysis of hyperspectral data. Furthermore, we develop an adaptive model to characterize the statistical distribution of the data. The model is automatically constructed according to the given data, which can encourage the algorithm to be appropriate for various hyperspectral images with different statistical characteristics. When applied to the hyperspectral data, the proposed approach can effectively overcome the problem of ICA-based methods, producing more accurate results. Even when the number of endmembers is incorrectly given, the method can still obtain desired results. As there is no need of spectral prior knowledge, this method provides an effective technique for the blind unmixing of hyperspectral imagery.2. A novel framework for endmember extraction is proposed. This is a geometrical method, while it also based on the algebra principle of triangular factorization (TF). By utilizing TF to calculate and analyze simplex volume, it requires just one comparison through the data to succeed in finding the global optimal solution for all the endmembers, thus improving the searching efficiency. This method and can obtain accurate results with very fast computational speed in high dimensional data, which can be applied in a realtime system. Besides, the dimensionality reduction transformation is not necessary for the proposed method, so the users can choose to reduce the dimensionality whether or not, depending on their practical situation. More importantly, since TF is a broad conception including different methods, the method is a framework including various implementations. From this framework, we can deduce different endmember extraction algorithms, and each algorithm has its own property. Theoritically, one can tell the the differences of these methods in either computational speed or numerical calculation, which has important meaning in the study of endmember extraction.3. Based on TF, the thesis also proposes an algorithm for abundance quantification. This is a fast learning method which can cooperate with the proposed TF-based endmember extraction. Besides estimating abundances, the method can also rectify possible errors in the given endmembers by utilizing important physical constraints of the hyperspctral mixture model. In this sense, the proposed method is useful for the imagery without pure pixels, and it improves the accuracy of geometrical methods.4. A method for unsupervised band selection is presented, for the purpose of dimensional reduciton. We introduce a new technique, complex networks, into the analysis of hyperspectral data. By ultilizng the networks’topological feature, we can evaluate the statistical characteristics and intrinsic properties of hyperspectral imagery, finding out the useful bands. The method searches for the network set which is most qualified for demarcating and identifying different substance signatures, and then the network set’s corresponding bands are regarded as the descried output results. Experimental results demonstrate that, the proposed method can reduce the dimensionality of data while preserving useful information, acquiring satisfactory results when applied for hyperspectral classification.