.png){kind=icon}
.png){kind=icon}
Return
Scale-Aware Mamba-Driven Implicit Representation for Arbitrary-Resolution Hyperspectral Pan-Sharpening
.png){kind=icon}
PRE
AI
DOI：10.1109/tgrs.2026.3707972.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Pan-sharpening aims to reconstruct high-resolution hyperspectral images (HR-HSI) by fusing the spectral information of hyperspectral images (HSI) with the spatial details provided by panchromatic (PAN) images. However, most existing pan-sharpening methods regard the fusion of hyperspectral and PAN images at different spatial resolutions as independent tasks, necessitating the training of a distinct model for each scale. This paradigm is impractical in real-world applications, as maintaining separate models for every spatial scale incurs prohibitive computational and storage costs in arbitrary-resolution pan-sharpening scenarios. Therefore, we propose the scale-aware Mamba-driven implicit representation network (SAMIR-Net), a unified scale-aware framework for arbitrary-resolution hyperspectral pan-sharpening. The proposed SAMIR-Net consists of two main components: scale-aware modality interactive extraction and implicit spatial–spectral reconstruction. The scale-aware modality interactive extraction is realized through a dual-branch, scale-sensitive subnetwork that collaboratively extracts scale-aware multimodal features and facilitates spatial–spectral interaction. The implicit spatial–spectral reconstruction module (IS2RM) further performs cross-attention between spatial and spectral embeddings, enabling flexible and robust coordinate-based pan-sharpening at arbitrary resolutions. Extensive experiments show that SAMIR-Net achieves favorable performance compared with representative state-of-the-art methods and exhibits improved generalization across different upsampling factors, indicating its potential for flexible hyperspectral reconstruction. The code is available at: https://github.com/Jiahuiqu/SAMIR-Net
Keywords:
Arbitrary-resolution reconstruction
hyperspectral pan-sharpening
implicit neural representation (INR)
Mamba
scale-aware learning
spatial–spectral fusion
Journal
IF:
8.6
Papers:
2.1W
Citations:
10.7W