GFHMP: Gradual Fusion Framework of Hyperspectral, Multispectral, and Panchromatic Images Using a Novel Spatial-Spectral Cross-Fusion Network

This article proposes a Gradual Fusion Framework for Hyperspectral (LR-HSI, 30 m), Multispectral (MR-MSI, 10 m), and Panchromatic (HR-PAN, 2.5 m) Images (GFHMP) to achieve large-scale multi-resolution satellite image fusion. The GFHMP framework employs an innovative Spatial-Spectral Cross-Fusion Network (SSICFNet) that effectively integrates complementary spatial and spectral information from multi-resolution inputs. The proposed SSICFNet utilizes a Spectrum Cross-Attention Block (SECB) and a Spatial Cross-Attention Block (SACB) to cross-fuse LR-HSI with MR-MSI (or MR-HSI with HR-PAN) images, effectively integrating spatial-spectral information for high-quality image reconstruction. The performance of GFHMP and SSICFNet was evaluated using both the public simulated Chikusei dataset and a custom real-world M-ZY1E dataset. Experimental results demonstrate that SSICFNet achieves state-of-the-art performance, outperforming existing methods on both datasets. Quantitative evaluation on the Chikusei dataset shows PSNR values of 43.94 dB (MR-HSI) and 42.94 dB (HR-HSI) for images fused with different spatial resolutions. The reconstructed vegetation spectral curve in HR-HSI exhibits exceptional agreement (R² = 0.9993) with the reference data. For the M-ZY1E dataset, the reconstructed images display high spatial consistency with panchromatic references and strong spectral correlation with original data, confirming the robustness of the GFHMP framework. To compensate for the lack of systematic validation regarding fused image feasibility in quantitative inversion, we employed LR-HSI, MR-HSI, and HR-HSI data from the M-ZY1E dataset to invert canopy chlorophyll content (CCC) in the Qinzhou Coastal Mangrove Reserve, Guangxi. The fused image achieves an inversion accuracy of R² = 0.759, representing an 11.49%-41.60% improvement over LR-HSI standalone inversion. We employed Shapley Additive Explanations (SHAP) to quantitatively assess the impacts of enhanced spectral fidelity and spatial resolution on chlorophyll retrieval. SHAP analysis demonstrates that HR-HSI texture features significantly improve mangrove CCC estimation, validating the advantages of image fusion for quantitative inversion. The source code is publicly available at: https://github.com/RSFubolin/GFHMP.