PMHIF-Net: A Prior-Guided Mamba Hierarchical Interactive Fusion Network for Hyperspectral and Multispectral Image Fusion

Abstract

Hyperspectral and multispectral image fusion (HMIF) aims to generate hyperspectral images (HSIs) with high-spatial resolution and favorable spectral fidelity. Benefiting from efficient and powerful global modeling capacity, Mamba has become a prevailing solution for the HMIF task in recent years. Nevertheless, most existing Mamba-based fusion methods fail to effectively capture inherent characteristics, including spatial edge structures and spectral reflectance variation rates during state-space modeling, and suffer from redundant information in the fusion pipeline. To address these limitations, this article proposes a prior-guided hierarchical interactive Mamba fusion network termed PMHIF-Net. First, edge priors and spectral gradient priors are embedded into the state-space model (SSM), and the spatial learning module (SpaLM) and spectral learning module (SpeLM) are developed. The priors explicitly participate in state-space evolution to capture local spatial edge details and spectral variation patterns across adjacent bands. Second, to correlate prior-enhanced features and enable progressive optimization of multilevel features, the spatial–spectral interactive Mamba block (SSIMB) is constructed. Deep cross-modal interaction is realized via bidirectional dynamic interaction and mutual control mechanisms. Finally, the refined gated reconstruction module (RGRM) is designed to couple and refine hierarchical features. Frequency-domain priors are introduced to build state-space gating, which retains valid high-frequency and low-frequency components and suppresses fusion redundancy. Quantitative and qualitative experiments are conducted on public datasets, including Pavia University, Houston, and Chikusei, as well as the self-built nonhomologous real Huzhou dataset. Experimental results demonstrate that the proposed PMHIF-Net outperforms the other state-of-the-art methods and achieves satisfying generalization and robustness.