Multitask Multiobjective Optimization Method for Adaptive Band Selection

Band selection is a crucial issue in hyperspectral image analysis. Due to variations in the data characteristics of hyperspectral images, it is necessary to perform adaptive band selection for each specific image. Current adaptive band selection methods primarily focus on the information content, redundancy, and number of bands based on unlabeled samples. However, these methods overlook the fact that utilizing a small number of labeled samples can greatly aid in selecting highly discriminative band subsets. This paper proposes a multitask multiobjective optimization method that simultaneously utilizes the discriminatory power of labeled samples and the information from abundant unlabeled samples to adaptively select the most suitable band subset. One task evaluates the information and redundancy of the selected band subset using unlabeled samples, while the other task incorporates labeled samples to assess the number of bands and discrimination of the selected subset. Experimental results demonstrate that the proposed adaptive band selection method effectively chooses the optimal band subset for a wide range of hyperspectral images, achieving high discrimination, information content, and low redundancy.