Dual-Domain Synergistic Pansharpening Network With Region-Adaptive Frequency Convolution

Abstract

Pansharpening is a critical technique in remote sensing aimed at generating high-resolution multispectral (HRMS) images by fusing high-resolution panchromatic (PAN) images with low-resolution multispectral (LRMS) images. However, existing methods face challenges in frequency-domain feature extraction, as global analyses often neglect regional characteristics, while local patch-based approaches may compromise the structural integrity of the image. To address these issues, we propose a novel pansharpening network utilizing a dual-branch architecture to extract frequency-domain features from PAN and MS images. This approach effectively harnesses their complementary information to enhance pansharpening performance. The extracted features are integrated with spatial-domain details via a hierarchical fusion (HF) module, enabling comprehensive reconstruction of HRMS images. In addition, we introduce a novel frequency-domain feature extraction method, termed region-based self-similarity adaptive frequency convolution (RSAFC). This method dynamically adjusts the frequency characteristics of distinct image regions by leveraging cluster-based self-similarity relationships and adaptive convolution operations that combine amplitude and phase, thereby achieving precise modeling of frequency-domain information. Experimental evaluations on the WorldView-3 (WV3) and QuickBird (QB) datasets demonstrate that the proposed method outperforms state-of-the-art approaches in both subjective and objective metrics.