Sparse Hyperspectral Unmixing With Preference-Based Evolutionary Multiobjective Multitasking Optimization

The traditional sparse unmixing methods based on multiobjective evolutionary algorithms (MOEAs) only deal with a single mixed pixel, without considering the spatial structure relationship between different mixed pixels. In addition, these methods suffer from the curse of dimensionality caused by the large number of pixels in hyperspectral image and spectra in library. In this paper, an evolutionary multitasking unmixing based on weakly nondominated sorting (EMTU-WNS) algorithm is proposed to alleviate these existing issues. First, a hyperspectral image is classified into multiple homogeneous regions, and the unmixing of pixels in the same region is constructed as a multiobjective optimization task. Then all the tasks are optimized simultaneously by using a population in the design of genetic transfer of intra-task and inter-task. In comparison with the original unmixing task with all pixels, these tasks in multiple homogeneous regions are relatively simple in term of dimensionality. Furthermore, it is inefficient for individuals to explore the whole search space. Therefore sparsity-constrained genetic operators are designed to evolve individuals towards the preference sparsity region. Finally, a preference-based weakly nondominated sorting is proposed to increase the number of nondominated solutions and maintain the diversity. The experimental results on three hyperspectral data sets demonstrate the effectiveness of EMTU-WNS with better convergence characteristics and unmixing accuracy.