RPM: Robust PAN-Sharpening With Multispectral Misalignment Correction

Most existing PAN-sharpening methods assume perfect co-registration between the panchromatic (PAN) image and multispectral (MS) bands. In real-world aerial and satellite imaging, however, both PAN–MS misalignment and inter-band misalignment among MS channels frequently occur, leading to significant degradation in fusion quality. In this work, we propose RPM, a robust PAN-sharpening framework that simultaneously corrects inter-band shifts while fusing PAN and MS information. The core idea is ashift-to-align training strategy, where random band-wise shifts are introduced during training so that the network learns alignment-invariant spectral fusion guided by the PAN image. Without requiring explicitly aligned ground truth, RPM enforces spectral fidelity, spatial consistency, and robustness to inter-band misalignment. Extensive experiments on multiple benchmark datasets demonstrate that RPM consistently achieves state-of-the-art performance: on the WV3 dataset, it reduces ERGAS by up to 2–3%, lowers SAM by about 1%, and improves PSNR, SCC, HQNR, and Q8 scores by 0.1–0.2% compared with recent methods. Moreover, RPM performs inference in the millisecond range per image patch, which is over 1000× faster than prior approaches that typically require several seconds.