Answers are grounded in this paper and cite their sources.
Hyperspectral image change detection (HSI-CD) has attracted increasing attention because hyperspectral images (HSIs) contain rich spectral information, which enables more detailed and accurate analysis of land cover change. To mitigate the influence of mixed HSI pixels caused by low spatial resolution, unmixing has been used in some HSI-CD methods. However, these methods suffer from the following limitations: first, most unmixing-based methods fail to fully exploit the temporal information contained in bitemporal HSIs, second, dimensional inconsistency between HSIs and abundance maps, which hinders effective pixel–subpixel feature fusion, and, third, lack of effective feature alignment to enhance complementary information and suppress irrelevant signals. To address these limitations, we propose a subpixel-guided network for HSI-CD based on multitemporal unmixing, called SGNet, which adaptively injects subpixel information into pixel features to enhance HSI-CD performance. In SGNet, a novel multitemporal unmixing block is designed to extract abundance matrices using shared endmembers, enabling accurate subpixel representation. Then, the spectral transformer captures pixel-level features with spatial–spectral change information. Finally, the spectral–spatial features of HSI and abundance features are adaptively fused by pixel–subpixel fusion module after feature alignment, enhancing discriminative representation for change detection. Experimental results on three HSI-CD datasets demonstrate that the proposed SGNet can achieve outstanding performance over existing state-of-the-art methods.
No takes yet. Share an insight, caveat, or question.
Dong et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: