S$^{3}$Mamba: Pan-Sharpening via Spatial–Spectral Synergistic State Space Model
Yuyan Yan, Yu Wang, Wei Tu, Jiaming Wang, Bowen Cai, Qingwei Zhuang, Xiaolong Zuo, Yunong Chen, Haobin Zhang, Zhenfeng Shao · IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing · 2026
Pan-sharpening aims to fuse a high-spatial-resolution panchromatic image with a low-spatial-resolution multispectral (MS) image to generate a high-quality, high-spatial-resolution MS image. Despite the development of existing pan-sharpening methods, they commonly face a critical challenge: insufficient modeling of deep-guided fusion mechanisms between spectral and spatial information, which hinders precise synergy and complementarity. To address this pivotal issue, this article proposes a spatial–spectral synergistic Mamba network (S $^{3}$ Mamba). S $^{3}$ Mamba adopts a dual-branch guided structure, designed to balance spectral consistency and spatial detail preservation effectively. Specifically, the network introduces several key innovations: in the initial feature extraction stage, dedicated spectral Mamba modules and spatial Mamba modules are designed to leverage the Mamba architecture for deeply modeling interchannel dependencies within the spectral dimension and for effectively capturing long-range spatial contextual information by applying Mamba along the spatial dimensions, respectively. Subsequently, during the feature fusion stage, we introduce a synergistic Mamba fusion module that processes features from different modalities using independent Mamba encoders and integrates a novel cross-modal feature guidance mechanism to achieve deep interconnection between spectral and spatial information. Furthermore, the network incorporates multimodal deep Mamba fusion modules that employ an early fusion strategy, ensuring that the fused features accurately represent both spectral and spatial information while also capturing multiscale details through stacked fusion blocks and local convolutional refinement. Extensive quantitative and qualitative experimental results on three different datasets demonstrate that the proposed S $^{3}$ Mamba method significantly outperforms existing state-of-the-art pan-sharpening techniques in terms of both objective evaluation metrics and visual perception quality.