Study Of Hyperspectral Sparse Unmixing Based On Evolutionary Multi-objective Optimization

Science and technology is contemporarily undergoing a major development among which hyperspectral sensing technology is noteworthy. The image data obtained by this technology has proved to be valuable and beneficial in many fields, such as environment pollution detection and geological prospecting. Generally, the technical limitations of sensor technology and diverse surface features lead to mixed pixels in image data. The mixed pixels must be unmixed to enable the further applications of the hyperspectral sensing technology. Pixel unmixing has become a matter of concern and aroused the interest of increasing number of scholars. Among the brunches of this technology, the sparse regression-based hyperspectral unmixing is the most wide-spread. It is used to find the optimal linear combination that expresses the measured spectrum of each mixed pixel using available spectrum libraries. The conventional sparse hyperspectral unmixing algorithm solves the problem by replacing the non-convex L0-norm with the convex L1-norm and then introducing the weight coefficient to integrate the two or more cost functions into one. However, this approach is not flawless.On one hand, the L1-norm is not always equivalent to L0-norm. On the other hand, the proper weight coefficient to be introduced is hard to obtain. To solve the problem mentioned above, the thesis performed the following studies:Firstly, this thesis presents an overview of the hyperspectral unmixing technology. The Hyperspectral unmixing model is usually divided into two parts based on the scale: linear and non-linear model. The linear model is the most widely used hyperspectral unmixing model and this thesis is mainly based on it. For linear model, signal-subspace, geometrical,statistical, sparse regression-based, and spatial-contextual unmixing algorithms are currently used.Secondly, this thesis builds a evolutionary multi-objective optimization model based on sparse regression-based hyperspectral unmixing algorithm to solve the problem. Besides, the multi-objective cooperative coevolution algorithm MOSU is proposed to solve the largescale optimization problem. Owing to the sparse nature of the fractional abundance of each mixed pixel, a sparse regression-based grouping strategy is put forward to better optimize the objective function. Furthermore, a knee point-based coordination mechanism is used to avoid complex computation. The simulation experiments demonstrate the effectiveness of the multi-objective sparse hyperspectral unmixing algorithm.Finally, the thesis add an objective function to efficiently utilize the spacial contextual information, thereby integrating the spectral information in the image and the spacial contextual information and bring the relationship between each pixel and the neighboring ones. Simulation experiments demonstrate that introducing spacial contextual information enhances the algorithm robust against noises.