A Robust Frequency Domain-Generalized Network for Multispectral Pansharpening

Multispectral pansharpening combines high-resolution panchromatic images and low-resolution multispectral images to generate high-resolution multispectral images. Existing deep learning-based pansharpening methods have been shown to outperform traditional ones by better preserving spatial and spectral details. However, most existing deep learning-based methods rely heavily on large amounts of training data from a single satellite, and their performance deteriorates significantly when applied to testing data from different satellites. The generalization problem of pansharpening methods is a challenge due to the shift of data distribution from training to testing data. A domain-generalized approach capable of handling cross-satellite scenarios would greatly reduce training costs and enhance practical applicability. Thus, we proposed a novel robust frequency domain-generalized pansharpening network (RFDGNet). RFDGNet incorporates frequency decomposition to better capture domain invariant characteristics, along with an adversarial training strategy. Specifically, RFDGNet consists of three specialized modules: the global dynamic spectral extractor, which dynamically learns satellite-specific styles based on low-frequency global information; the gradient-adaptive high-frequency enhancer, which refines spatial details through gradient analysis and wavelet decomposition; and the spectral enhancement module for spectral information preservation. Furthermore, we design a novel domain generalization strategy, which includes a new adversarial loss to improve the network and employ misleading domain labels to guide the network in learning domain-invariant representations. Experiments on both in-domain and cross-domain pansharpening scenarios show significant advantages in both accuracy and generalization.