Physics-guided supervision for highly generalisable deep transformer fusion of hyperspectral and multispectral data: application to Sentinel-2 and EnMAP

Fusing hyperspectral (HS) and multispectral (MS) satellite imagery enables the joint exploitation of fine spatial detail and rich spectral information, supporting accurate environmental and geological analyses. Deep Learning methods have recently become the top-performing solutions for HS–MS fusion, yet their progress is limited by the scarcity of large, diverse training datasets. Existing datasets, mainly derived from airborne acquisitions, cover small regions and lead to poor generalisation and network overfitting. To address this, we introduce a physics-based framework that generates realistic HS–MS training data directly from satellite sensors. Applied to Sentinel-2 (S2) and EnMAP, it combines S2’s spatial detail with EnMAP’s spectral signatures while simulating each instrument’s Point Spread Function, Modulation Transfer Function, and Spectral Response Function. This approach enables the creation of large, physically consistent, globally distributed datasets with strong generalisation capability. A transformer-based autoencoder trained on these synthetic data achieves high spatial and spectral fidelity in S2–EnMAP fusion. Comparable performance obtained with other Deep Learning architectures confirms that the physics-grounded supervision, rather than the network design, is the main driver of success. Validation on real Sentinel-2 and EnMAP imagery demonstrates the robustness of this scalable, sensor-specific framework for practical, global HS–MS fusion.