Physics-Guided Generative Pansharpening for Remote Sensing Images Under Unknown Sensor Degradation

Abstract

Generative artificial intelligence has shown considerable potential for remote sensing image restoration and fusion. Pansharpening reconstructs high-spatial-resolution multispectral (HRMS) images by integrating high-resolution panchromatic (PAN) and low-resolution multispectral (LRMS) observations. However, existing methods often depend on reduced-resolution supervision and fixed degradation models, whereas real satellite imagery exhibits unknown sensor- and scene-dependent spatial degradation, modulation transfer functions, and PAN spectral responses. Most methods also yield deterministic reconstructions without quantifying uncertainty. This article proposes a physics-guided generative pansharpening framework that formulates reconstruction as posterior inference with latent spatial degradation and spectral response parameters. A lightweight self-calibration module estimates the blur kernel and PAN spectral response from observed image pairs. A residual posterior diffusion model then generates HRMS samples conditioned on the calibrated imaging model, while observation-guided reverse sampling enforces consistency with both PAN and LRMS observations. Multiple posterior samples are used to obtain a fused image and an uncertainty map reflecting reconstruction reliability. Experiments on QuickBird, GaoFen-2, and WorldView-2 datasets under reduced- and full-resolution evaluations, scene-level cross-validation, and spectrally compatible cross-sensor settings demonstrate improvements in spectral preservation, spatial detail reconstruction, and target-sensor adaptation over traditional, convolutional neural network-based, generative adversarial network-based, and diffusion-based methods. Ablation studies further confirm the effectiveness of selfcalibration, observation-guided generation, and posterior uncertainty estimation. These results highlight the potential of physics-guided generative learning for reliable and interpretable remote sensing image fusion.