Pansharpening by Convolutional Neural Networks in the Full Resolution Framework

In recent years, there has been a growing interest in deep learning-based pansharpening. Thus far, research mainly focused on architectures. Nonetheless, model training is an equally important issue. A first problem the absence of ground truths, unavoidable This often addressed by networks reduced resolution domain and using original data as truth, relying implicit scale invariance assumption. However, full images results are disappointing, suggesting such not to hold. further scarcity data, which causes limited generalization ability poor performance off-training test images. this paper, we propose full-resolution framework for The fully general can be used any pansharpening model. Training takes place high-resolution domain, only thus avoiding loss information. To ensure spectral spatial fidelity, suitable two-component defined. component enforces consistency between pansharpened output low-resolution multispectral input. component, computed at high-resolution, maximizes local correlation each band panchromatic At testing time, target-adaptive operating modality adopted, achieving good with computational overhead. Experiments carried out WorldView-3, WorldView-2, GeoEye-1 show that methods trained proposed guarantee pretty terms both numerical indexes visual quality.