DPP: Scale-Generalization Transformer Based on Dynamic Projection for Pansharpening

This letter introduces a scale-generalization Transformer for pansharpening (DPP), ensuring high performance across resolutions. First, we employ joint learning of the pansharpening model across resolutions, establishing a collaborative iteration mechanism between dynamic full-resolution (FR) spectral projection model estimation and the pansharpening process. This effectively reduces errors introduced by directly using the reduced-resolution spectral projection model, guiding the pansharpening model to generalize from the well-defined reduced-resolution scale to the ill-posed F scale. Second, we customize a novel network for pansharpening, employing inverse spatial injection to enhance local and global awareness. Initially, convolutional neural network (CNN) captures detailed spatial features from the panchromatic image, which are then inversely downsampled for aligned injection. Subsequently, CNN and Transformer are combined to integrate global spectral features with local spatial features, ensuring high-quality pansharpening. Extensive experiments demonstrate the advantages of our DPP in terms of information fidelity at both reduced and full resolutions.