STG-Net: Shared Tensor Guidance Hyperspectral and Multispectral Remote Sensing Image Fusion Network Under Locally Overlapping Condition

Hyperspectral image (HSI) and multispectral image (MSI) fusion aims to generate a high spatial and spectral resolution hyperspectral image (HR-HSI) by integrating the spectral information of low-resolution HSI (LR-HSI) and the spatial information of high-resolution MSI. However, most existing fusion methods are developed under the fully overlapping assumption, and their ability to exploit locally shared information becomes limited when the two modalities only cover a partially common area. In addition, directly extending the tensor representation learned from the overlapping region to nonoverlapping regions is challenging because these regions lack LR-HSI spectral observations. To overcome these issues, in this article, an unsupervised shared tensor guidance network, termed STG-Net, is proposed for HSI-MSI fusion under locally overlapping conditions. Specifically, a local-to-global deep Tucker representation is constructed to learn a shared core tensor from the reliable overlapping area. Then, a multiscale attentive core synthesizer is designed to propagate the learned core representation to the full scene by exploring multiscale spatial-spectral features. Finally, an adaptive weighted fusion reconstructor reconstructs the HR-HSI by adaptively combining the overlap-constrained reconstruction and the full-scene MSI-guided reconstruction. The whole framework is optimized without ground-truth HR-HSI by using observation-domain degradation consistency, overlap-region cross-reconstruction, style consistency, and manifold regularization. Experiments on three benchmark datasets under ×4 and ×8 downsampling settings demonstrate the effectiveness of STG-Net. On the Pavia University dataset under ×4 downsampling and 80% overlap, STG-Net achieves a peak signal-to-noise ratio of 42.50 dB.