EvoAR-SFNet: Evolutionary Autoregressive Spatial-Frequency Network for Remote Sensing Pansharpening

Pansharpening aims to generate high-resolution multispectral (HRMS) images by fusing high-resolution panchro matic (PAN) images with low-resolution multispectral (LRMS) inputs. However, existing methods often face difficulties in effectively using spatial and frequency information, mainly due to limited fusion strategies and the lack of progressive refinement mechanisms. To overcome these limitations, we propose EvoAR SFNet (Evolutionary Autoregressive Spatial-Frequency Network), a novel framework for remote sensing pansharpening, comprising two specialized branches: a spatial branch employing Adaptive Rectangular Convolutions (ARConv) to capture spatial details, and a frequency branch adopting Frequency-Dynamic Convolu tions (FDConv) to extract frequency-aware features. Within each branch, Evolutionary Autoregressive (EvoAR) module performs multi-stage progressive refinement using autoregressive feedback. In addition, it adaptively adjusts loss weights using evolutionary optimization. The refined features from both branches are sub sequently integrated using a Bi-Domain Fusion Block (BDFB), which enhances cross-domain feature interactions through im plicit modulation, producing high-fidelity high-resolution mul tispectral outputs. Extensive experiments on WV3, QB, and GF2 datasets demonstrate that EvoAR-SFNet achieves competitive overall performance, especially in full-resolution evaluation, where it obtains strong spatial-detail preservation and leading QNR results.