Block Term Decomposition-Guided Frequency Mamba Modulation for Hyperspectral Image Fusion

Owing to the complementary spatial-spectral characteristics of hyperspectral images (HSIs) and multispectral images (MSIs), HSI-MSI fusion has become an effective approach for generating high-resolution hyperspectral images (HR-HSIs). Recent deep learning frameworks, particularly Mamba-based architectures, provide efficient long-range dependency modeling with near-linear complexity. Nevertheless, they often overlook the intrinsic low-rank structure of hyperspectral data and are largely insensitive to frequency-domain information. To address these limitations, we propose block term decomposition-guided frequency Mamba modulation (BFMM), a low-rank tensor modulation framework built upon block term decomposition (BTD). BFMM represents features as a sum of multiple Tucker components, each reconstructed by the multimode product of a learnable Tucker core and low-rank factor matrices, thereby preserving structured spatial-spectral dependencies. A frequency Mamba joint modulation (FMJM) unit learns frequency-aware low-rank Tucker cores through adaptive frequency conditioning, and a Kronecker-rank interaction modulator (KRIM) learns spatial-spectral factor matrices via separable channel-rank guidance in the form of Kronecker products. The modulated core tensors are sequentially contracted with the learned factors through mode-n products, thereby reconstructing low-rank feature tensors whose aggregation yields structurally interpretable and information-dense representations. Across three public datasets, BFMM achieves state-of-the-art reconstruction performance with a favorable accuracy-complexity trade-off, and is further validated on the WHU-MHF and WV-3 datasets. The source code is available at https://github.com/lirui818/BFMM.