Shared Tensor Guidance Super-resolution Network for Locally Overlapping Hyperspectral and Multispectral Remote Sensing Image Fusion

Hyperspectral and multispectral image fusion aims to generate a high-resolution hyperspectral image (HR-HSI) from a low-resolution hyperspectral image (LR-HSI) and a high-resolution multispectral image (HR-MSI). However, existing methods often suffer from performance degradation in non-overlapping regions - i.e., areas covered only by HR-MSI - resulting in the loss of fine spatial details and spectral consistency. To address this issue, we propose a novel fusion method named STG-Net (Shared Tensor Guidance Super-resolution Network), based on core tensor guidance and enhanced super-resolution for non-overlapping regions. First, a deep Tucker decomposition network is employed to learn a shared core tensor and corresponding factor matrices from the overlapping regions. Second, a Multiscale Attentive Core Synthesizer (MACS), incorporating multi-scale residual blocks, dense channel attention, and spectral-spatial co-attention, is designed to generate a full-scene core representation under the guidance of the overlapping core tensor. Finally, an Adaptive Weighted Fusion Reconstructor (AWFR) is introduced to seamlessly integrate overlapping and non-overlapping regions. Experimental results on the Pavia University dataset under 4× downsampling demonstrate that the proposed STG-Net significantly outperforms existing approaches, effectively eliminating grid artifacts while preserving spectral fidelity.