Domain-Consistent Pansharpening via Learnable Degradation Modeling and Amplitude-Phase Modulation

Deep learning-based pansharpening methods have achieved promising results under Wald’s protocol. However, their effectiveness on real satellite imagery is constrained by discrepancies between simulated degradation and real observations. Moreover, existing frequency-domain fusion methods insufficiently exploit the distinct yet complementary roles of amplitude and phase information. To address these issues, we propose a novel Domain-Consistent Pansharpening framework via Learnable Degradation Modeling and Amplitude–Phase Modulation (DC-APM). First, separate degradation networks are trained for multispectral (MS) and panchromatic (PAN) images to generate data-adaptive training samples through reconstruction-constrained local distribution alignment, reducing reliance on predefined degradation assumptions. Second, we design an Amplitude–Phase Modulation Network (APM-Net) for feature fusion. An Amplitude–Phase Modulation Module (APMM) enhances spectral and structural representations through explicit amplitude–phase decomposition, while a lightweight Cross-domain Interaction Unit (CIU) enables bidirectional information exchange for adaptive frequency-domain modulation. Finally, a progressive multi-scale reconstruction strategy integrates the modulated features to generate high-resolution MS images. Extensive experiments on four satellite datasets demonstrate the effectiveness of the proposed framework under both reduced-resolution and full-resolution evaluation protocols, achieving competitive or superior performance compared with representative state-of-the-art methods.