Evolutionary Multitasking With Adaptive Cross-Dataset Knowledge Transfer for Band Selection of Hyperspectral Images

Band selection (BS) is a critical method for hyperspectral image (HSI) classification, which helps to reduce the computational burden while providing a good class separability. However, most BS methods only deal with each HSI dataset independently, which cannot effectively exploit useful knowledge across multiple datasets. Only one BS method tries to transfer knowledge across multiple datasets, but it ignores the negative transfer effect caused by the inherent difference among datasets. To alleviate this issue, this paper proposes an evolutionary multitasking BS method with adaptive cross-dataset knowledge transfer. Specifically, a new knowledge transfer strategy is designed based on the linear mapping to filter out irrelevant knowledge among different datasets. Then, an adaptive transfer control strategy is designed to adaptively adjust the frequency and intensity of knowledge transfer, which can further enhance the knowledge transfer performance. As validated by our experimental studies, our method can properly mitigate the negative transfer effect caused by the differences in cross-dataset knowledge transfer. When compared to several state-of-the-art BS methods on three common HSI datasets, our method can find superior band subsets with higher quality.