Multiview representation-guided global-local fusion for hyperspectral image change detection

Despite the significant progress achieved by deep learning (DL) in hyperspectral image change detection (HSI-CD), several challenges still constrain its performance. In particular, the land cover complexity makes it difficult to accurately analyze temporal differences, and the boundaries between changed and unchanged regions are often difficult to distinguish. To address these issues, this paper proposes a multiview representation-guided global-local fusion method (MVR-GLF) for HSI-CD, which consists of four key components: a multiview representation module (MVRM), a Mamba-based global difference feature extraction module (GDFEM), a local frequency-division feature extraction module (LFDFEM), and a bidirectional dynamic fusion module (BiDFM). Specifically, MVRM constructs a joint representation that integrates temporal spectral numerical differences and shape differences from bi-temporal HSIs, providing explicit guidance for subsequent modules to focus on temporal differences. GDFEM leverages Mamba to model long-range dependencies within the joint representation, enabling the extraction of global pixel-wise difference features. LFDFEM processes the low- and high-frequency components derived from discrete wavelet transform separately, enhancing boundary information between changed and unchanged regions. Finally, BiDFM generates multi-scale dynamic convolution kernels to facilitate bidirectional fusion of global pixel-wise difference features and local fine-grained details. Extensive experiments conducted on three benchmark datasets demonstrate that the proposed method consistently outperforms existing approaches. The source code is available at https://github.com/Preston-Dong/MVR-GLF.