Deep unrolling for unified pansharpening with satellite-specific and image-adaptive priors

Pansharpening is the task of fusing low-resolution (LR) multispectral (MS) and high-resolution (HR) panchromatic (PAN) images into HRMS outputs, which is critical for improving the spatial and spectral utility of satellite imagery. While deep learning (DL)-based methods have achieved impressive results, most are trained separately for each satellite dataset. This “one-model-per-satellite” paradigm severely limits practical deployment, whereas directly training a single model on mixed multi-sensor data often causes to severe feature entanglement and yields a model that lacks physical transparency. To tackle this challenge, we propose UP-PAN (Unified and Physically-principled Pansharpening), a novel deep unrolling (DU)-based framework designed for unified multi-satellite pansharpening. Our method models the unified pansharpening task under a Maximum a Posteriori (MAP) formulation, incorporating both spatial and spectral observation constraints. To prevent feature entanglement, we explicitly decouple the complex multi-sensor prior into a satellite-specific prior and an image-adaptive prior. Concurrently, the satellite-specific prior is encoded through a discrete codebook mechanism to capture unique high-quality (HQ) priors from diverse HRMS data, while the image-adaptive prior utilizes a proximal network guided by spectral and spatial prompts to refine continuous spatial details. We conduct extensive experiments across four widely-used satellite datasets under a unified training protocol. Results demonstrate that a single, jointly trained UP-PAN model achieves state-of-the-art performance across all benchmarks, outperforming or matching satellite-specific models. The code is publicly available at .