PANDA: PAN Alignment-Aware Distillation via Adaptive Fitting for Pansharpening

Pansharpening aims to fuse high-resolution panchromatic (PAN) images with low-resolution multispectral (MS) images to reconstruct high-resolution multispectral images. However, satellite observations are often limited and highly sensor-dependent, while PAN and MS are acquired through different sensor paths, resolutions, and imaging geometries, leaving residual spatial displacement. Directly fitting such data can therefore reinforce incorrect spatial correspondence. To address this, we propose an align-first, fit-better framework for data-specific pansharpening. First, a PAN aligner is jointly optimized with a reconstruction network through HRMS reconstruction and self-supervised shift consistency, enabling task-relevant alignment without displacement labels. Then, the fixed reference model guides alignment-aware distillation, where reconstruction error balances hard/soft supervision and alignment-consistency error controls reliable spatial-detail fitting. Extensive experiments show improved PSNR and HQNR with fewer parameters, demonstrating robust spatial-spectral reconstruction through alignment-aware data fitting.