Multiobjective Optimization Based Subpixel Mapping For Hyperspectral Remote Sensing Imagery

With the rapid development of hyperspectral remote sensing,land-cover interpretation using hyperspectral image has become an important means for national needs such as land-use mapping,environmental monitoring,resource investigation,and military reconnaissance.However,due to the limited spatial resolution of the hyperspectral data,and the complex and diverse distribution of land cover objects,mixed pixels are widely distributed in hyperspectral image.The mixed pixels with mixed spectrum hinder the development of quantitative hyperspectral remote sensing.Subpixel mapping is a useful technique to interpret the distribution of component objects within the mixed pixels.It uses the rich spectral information of hyperspectral images to make up for the lack of spatial information.The mixed spectrum is analyzed by spectral unmixing,which can identify the type and abundance of the subpixel targets in the mixed pixel.Based on the abundance map,subpixel mapping is applied with a variety of spatial distribution priori information,to infer the spatial distribution of the subpixel targets.Auxiliary data such as spatial distribution dictionary,multiple shifted images,time-sequence images,etc.can be used to provide such a spatial priori with geological complexity for subpixel mapping.However,with the introduction of a variety of prior information,subpixel mapping also has the following problems:(1)It is difficult to optimize the non-convex subpixel mapping model of single-image.Traditional subpixel mapping methods lack an intelligent optimization method.They convert the non-convex subpixel mapping model to a convex problem through convex relaxation;(2)It is difficult to effectively integrate the priori information of the auxiliary data to constrain the subpixel mapping problem.The traditional subpixel mapping methods use sensitive weights to combine the prior information into a single objective function to construct the mapping model.However,the determination of weight parameters is subjective.For different data and different scenes,they needs to be empirically adjusted,so it is difficult to achieve the optimal integration of the priori information.(3)Conventional subpixel mapping methods are developed based on the abundance maps produced by spectral unmixing.However,the unmixing and mapping processes are optimized independently,and the unmixing error will restrict the reliability of the subpixel mapping model.Therefore,to solved the aforementioned problems,this paper has carried out research on the multiobjective optimization based subpixel mapping methods for hyperspectral remote sensing imagery.The contents and contributions are as follows:(1)The ill-posed problem of subpixel mapping is systematically analyzed,and various priori information constraints introduced to make the subpixel mapping problem well-posed are introduced.The subpixel mapping algorithms system is summarized based on the various priori information constraints.(2)In-depth analysis of the potential of multiobjective optimization theory in solving the optimization problems of subpixel mapping,and sorted out the multiobjective evolutionary algorithm system and basic framework.(3)A sparse subpixel mapping method based on multiobjective optimization is proposed,which construct a multiobjective subpixel mapping model with non-convex sparse constraint,ensuring the modeling accuracy for subpixel mapping with singleimage.The multiobjective evolutionary algorithm with global optimization capability is applied to optimize the sparse coefficient solution of the non-convex mapping model,which improves the intelligent optimization ability of the subpixel mapping method.Experiments show that the overall accuracy of the proposed method is 2.5% higher than the optimal comparison algorithm.(4)A multiobjective spatiotemporal subpixel mapping method and a global-local search based subpixel mapping method with multiple shifted images are proposed,in which the priori information constraints of the auxiliary images are modeled as multiple objective functions for simultaneous optimization.The multiobjective evolutionary algorithm with multi-source information fusion ability enables the adaptive integration of the priori information constraints for subpixel mapping.The global-local search strategy is designed to enhances the optimization ability of the multiobjective evolutionary algorithm,to search for the subpixel spatial distribution map that achieves the optimal balance between various priori information constraints.Although the overall accuracy of the proposed method is only 1.6% higher than that of the optimal comparison algorithm,the sensitivity analysis shows that the proposed method is quite stable and the accuracy fluctuates within 0.3%.(5)With the framework of multiobjective optimization,a joint unmixing and subpixel mapping method is proposed,which is different from the conventional subpixel mapping procedure.Spectral unmixing and subpixel mapping are simultaneously optimized,which effectively improves the accuracy of unmixing and mapping,and reduces the dependence of subpixel mapping on the quality of abundance maps.Experiments show that the overall mapping accuracy of the proposed method is 3% higher than the optimal comparison algorithm,while the unmixing accuracy is also improved by 0.6d B.(6)A prototype system for subpixel mapping is developed based on the research of the four methods proposed in this thesis.The applicable scenarios and the basis for preferring one are analyzed by comprehensive comparison of the four methods.