FreDA: Frequency-Guided Degradation Adaptation Detail Injection Network for Pansharpening

Pansharpening fuses panchromatic (PAN) images with low-resolution multispectral (LRMS) images to generate high-resolution multispectral (HRMS) images with rich spatial and spectral information, which can be widely used in change detection and other downstream tasks. However, existing pansharpening networks often use fixed spatial degradation to construct training data. Moreover, conventional fusion paradigms in both spatial and frequency domains generally rely on predetermined strategies, which fail to dynamically adapt to the distinctive characteristics of individual images. To solve the above problems, we proposed a frequency-guided detail injection pansharpening network (FreDA), which consists of a spectral restriction sub-network (SimNet) and an image fusion sub-network (FNet). The SimNet uses the learnable convolution layer to characterize the degradation parameters, which is used to construct spectral restriction loss. The FNet is constructed with the proposed frequency-guided detail injection (FGI) module, which is designed based on the relationship between frequency information and image degradation condition. The proposed FreDA framework introduces a degradation-aware synchronization mechanism that dynamically aligns spatial degradation patterns with adaptive detail injection weights, achieving significant quality enhancement in multi-satellite data. Extensive experiments on both simulated and full-resolution datasets demonstrate its superiority over state-of-the-art methods. Furthermore, rigorous module-level ablation studies validate the indispensability of each component.